News Archive

Thesis presentations

Time: 22.05.2019, 10:00
Location: Room 072 in the Fraunhofer IGD, Fraunhoferstrasse 5, S3|05
Speaker: Moritz Sachs (Advisor: Matthias Unbescheiden)
Title: "Automatisierte Geschäftsmodellanalysen mit Deep Neural Networks" (Master thesis)
Abstract: Ein Hauptkriterium für das Investment eines Venture Capital Fonds in ein Start-up ist dessen Geschäftsmodell. Dieses ist im Businessplan enthalten. Das Screening, sowie die Analyse der eingereichten Businesspläne, erfolgt bei den meisten Venture Capital Fonds überwiegend durch Menschen.
Mit der vorliegenden Arbeit wird untersucht, inwieweit die Analyse der in den Businessplänen enthaltenen Geschäftsmodelle mit Hilfe von Deep Neural Networks automatisiert werden kann. Ziel war die Entwicklung eines Prototypen, der die in den Businessplänen enthaltenen Geschäftsmodelle automatisch extrahiert und in das Metamodell Startup Navigator überführt.
Dem Knowledge Discovery in Databases Prozess folgend wurden hierfür die Businesspläne eines Venture Capital Fonds aufbereitet und damit ein tiefes Convolutional Neural Network, der Multilabel k-Nearest Neighbour Algorithmus, sowie eine Support Vector Machine mit Naive Bayes Features trainiert.
Die Ergebnisse des entwickelten Prototypen zeigen, dass die in den Businessplänen enthaltenen Geschäftsmodelle automatisch extrahiert und in das Metamodell Startup Navigator überführt werden können. Es erscheint plausibel, dass mit mehr Trainingsdaten und einer intensiveren Hyperparameteroptimierung die Korrektklassifizierungsrate verbessert werden kann, sodass der Prototyp zum Aufbau eines Geschäftsmodellkorpus genutzt werden könnte.

Time: 30.04.2019, 10:00
Location: Room 103 in the Fraunhofer IGD, Fraunhoferstrasse 5, S3|05
Speaker: Moritz Matthiesen (Advisor: Pavel Rojtberg)
Title: "Interpolation von Kalibrierdaten für Zoom und Autofokus Kameras" (Bachelor thesis)
Abstract: In dieser Arbeit wird das Problem betrachtet, dass für jede neue Kameraeinstellung eine neue
Kalibrierung vorgenommen werden muss.
Ziel dabei ist Kalibrierdaten an bestimmten Kameraeinstellungen zu erstellen, um mithilfe von
diesen die Kalibrierdaten von anderen Kameraeinstellungen herzuleiten. Dabei werden die Kalib-
rierdaten betrachtet und es wird versucht Beziehungen zwischen den einzelnen Parametern der
Kalibrierung herzuleiten. Um diese zu ermitteln wird zwischen verschiedenen Parametern der
Kalibrierung interpoliert.

Time: 29.04.2019, 15:00
Location: Room 324 in the Fraunhofer IGD, Fraunhoferstrasse 5, S3|05
Speaker: Ali Jabhe (Advisor: David Kügler)
Title: "Physical World Attacks in Medical Imaging" (Bachelor thesis)
Abstract: The methodology and the acquisition of images for the attacks on dermoscopy tackles the question of whether Deep-Learning Systems can be attacked by a malicious attacker without changing anything on the Deep-Learning side. That mean only changes on the physical world are allowed. This problem is an extension of the "adversarial attack" concept, but with some twist.

Time: 25.04.2019, 14:00
Location: Room 324 in the Fraunhofer IGD, Fraunhoferstrasse 5, S3|05
Speaker: Heiko Reinemuth (Advisor: Juergen Bernard)
Title: "Visual-Interactive Labeling of Multivariate Time Series to Support Semi-Supervised Machine Learning" (Master thesis)
Abstract: Labeling of multivariate time series is an essential requirement of the data-centric decision-making processes in many time-oriented application domains. The basic idea of labeling is to assign (semantic) meaning to specific sections or time steps of the time series and to the time series as a whole, accordingly. Hence, weather phenomena can be  characterized, EEG signals can be studied, or movement patterns can be marked in sensor data.
In the context of this work a visual-interactive labeling tool was developed that allows non-expert users to assign semantic meaning to any multivariate time series in an effective and efficient way. Enabling experts as well as non-experts to label multivariate time series in a visual-interactive way has never been proposed in the information visualization and visual analytics research communities before. This thesis combines active learning methods, a visual analytics approach, and novel visual-interactive interfaces to achieve an intuitive data exploration and labeling process for users. Visual guidance based on data analysis and model-based predictions empowers users to select and label meaningful instances from the time series. As a result, the human-centered labeling process is enhanced by algorithmic support, leading to a semi-supervised labeling endeavor combining strengths of both humans and machines.

Time: 17.04.2019, 10:00
Location: Room 242 in the Fraunhofer IGD, Fraunhoferstrasse 5, S3|05
Speaker: Johann Reinhard (Advisor: Alan Brunton)
Title: "Efficient streaming sample-based surface triangulation of voxel data" (Master thesis)
Abstract: Voxel-based discrete representations of 3-dimensional data are widely used in several fields of graphical computing, for instance in the 3D printing driver Cuttlefish. For commonly used techniques, such as the marching cubes algorithm, the creation of a polygonal/polyhedral mesh representation of the used voxel data at high resolutions can become time-consuming and result in meshes with excessive numbers of vertices, which nonetheless introduce "staircase" artifacts relative to the desired geometry. It is then often necessary to use additional post-processing steps, such as mesh decimation, at the expense of additional computational effort and possible inaccuracies regarding the representation of the original shape.
The goal of this thesis is to simultaneously address all three of these issues, proposing an efficient technique to generate low-polygon meshes, which accurately represent the object’s shape. The intended technique is based on sampling the surface at regions of high curvature using, for example, an importance sampling technique, although different techniques will be explored. A comparison will be made between per-slice and per-chunk sampling (i.e. consider only a single slice or a whole chunk of slices when deciding where to place samples). The samples are to be mapped to a parametric, planar space, allowing to efficiently triangulate the sampled points. The necessity of additional post-processing steps in the parametric or reprojected object space will be assessed. The developed techniques will be implemented, integrated into Cuttlefish and evaluated based on comparisons to standard techniques such as marching cubes or Marching Tetrahedra using the above three measures: efficiency (time and memory), number of polygons in the output, and accuracy. Defining a measure of the accuracy of the output and computing it is a further aspect of the thesis, where at least the Hausdorff distance and collinearity of the surface normals will be measured in order to quantify the mesh quality.

Time: 20.03.2019, 15:00
Location: Room 073 in the Fraunhofer IGD, Fraunhoferstrasse 5, S3|05
Speaker: Alexander Distergoft (Advisor: Anirban Mukhopadhyay)
Title: "Interpreting Adversarial Examples in Medical Imaging" (Master thesis)
Abstract: Deep neural networks (DNNs) have been achieving high accuracy on many important tasks like image classification, detection or segmentation. Yet, recent discoveries have shown a high degree of susceptibility for these deep-learning algorithms under attack. DNNs seem to be vulnerable to small amounts of non-random noise, created by perturbing the input to output mapping of the network. These perturbations can severely affect the performance of DNNs and thus endanger systems where such models are employed.
The purpose of this thesis is to examine adversarial examples in clinical settings, be it digitally created or physical ones. For this reason we studied the performance of DNNs under the following three attack scenarios:
1. We hypothesize that adversarial examples might occur from incorrect mapping of the image space to the lower dimensional generation manifold. The hypothesis is tested by creating a proxy task of a pose estimation of surgical tools in its simplest form. For this we define a clear decision boundary. We use exhaustive search on a synthetic toy dataset to localize possible reasons of successful one-pixel-attacks in image space.
2. We design a small scale prospective evaluation on how Deep-learning (DL) dermoscopy systems perform under physical world attacks. The publicly available Physical Attacks on Dermoscopy Dataset (PADv1) is used for this evaluation. The introduced susceptibility and robustness values reveal that such attacks lead to accuracy loss across popular state-of-the-art DL-architectures.
3. As a pilot study to understand the vulnerabilities of DNNs that perform under regression tasks we design a set of auxiliary tasks that are used to create adversarial examples for non-classification-models. We train auxiliary networks on augmented datasets to satisfy the defined auxiliary tasks and create adversarial examples that might influence the decision of a regression model without knowing about the underlying system or hyperparameters.

Time: 18.02.2019, 15:00
Location: Room 072 in the Fraunhofer IGD, Fraunhoferstrasse 5, S3|05
Speaker: Fadi Boutros (Advisor: Naser Damer)
Title: "Reducing Ethnic Bias of Faces Recognition by Ethnic Augmentation" (Master thesis)
Abstract: Automated face recognition has gained wider deployment ground after the recent accuracy gains achieved by deep learning techniques. Despite the rapid performance gains, face recognition still suffers from very critical issues. One of the recently uncovered, and very sensitive, challenges is the ethnicity bias in face recognition decision. This is the case, unfortunately, even in the latest commercial and academic face recognition system. In 2018, the National Institute of Standards and Technology (NIST) published the latest report regarding the evaluation result of commercial face recognition solutions from several major biometric vendors. The report specifically evaluated and demonstrated the variance of the error rates of the evaluated solutions based on demographic variations.
This thesis is one of the first research efforts in addressing the decision bias challenge in biometrics. It builds its hypothesis on the strong assumption that ethnicity bias is caused by the relative under-representation of certain ethnicities in training databases. This work introduces a novel ethnicity-driven data augmentation approach to reduce biometric bias. The proposed approach successfully utilize a generative image model to generate new face images that preserve the identity of their source images while partially transforming their ethnicity to the targeted ethnicity group. A large-scale ethnicity-labeled face images database is created in order to develop and evaluate the proposed approach. To achieve that, part of this thesis focused on creating an ethnicity classifier to annotate face images, achieving accuracy in the state-of-the-art range. The proposed ethnicity-driven face generative model is developed based on the ethnicity labeled images to generate realistically and high-resolution face images, depending on a limit amount of training data. More importantly, the thesis proves that the proposed augmentation approach strongly preserves the identity of the input images and partially transforms the ethnicity.
The augmented images are used as part of the training data of a face recognition model. The achieved verification results prove that the proposed ethnicity augmentation methods significantly and consistently reduced the ethnicity bias of the face recognition model. For examples, the ERR was reduced from 0.159 to 0.130 when verifying inter-ethnicity samples of Black individuals on a model trained on Asian individual images, respectively before and after applying the proposed training data augmentation. Moreover, the overall performance of the face recognition model was improved. However, this improvement was more significant, as intended, in the targeted ethnicity groups.

Poster Presentation

Time: 13.02.2019, 10:00 - 11:00
Location: Room 073 in the Fraunhofer IGD, Fraunhoferstrasse 5, S3|05
Speaker: Students of the "Deep Learning for Medical Imaging"-course (Anirban Mukhopadhyay, David Kügler)
Topic: "Segmentation problems of medical images"

Thesis presentations

Time: 30.01.2019, 11:00
Location: Room 140 in the Fraunhofer IGD, Fraunhoferstrasse 5, S3|05
Speaker: Janina Hüther (Advisor: Helmi Ben Hmida)
Title: "Smart Recommendation Systems for IoT" (Master thesis)
Abstract: Nowadays Internet of Things is a important research field in Computer Science. The personalization of IoT services is important for the usability and user experience, but the rapidly growing number of IoT services and therefore the increasing number of possibilities, makes it hard for the user to configure an IoT system. Recommender systems are there to automate the decision process. But a lot of different recommender algorithms exist and it is hard to decide for a certain method regarding a specific IoT Use Case. In this master thesis recommender system methods are investigated in regard to their characteristics as well as IoT properties and scenarios. The goal of this research is to suggest a generic model for selecting and executing recommender system methods depending on a given IoT scenario. All defined tasks were accomplished successfully. A complete concept was created and realized as an web interface application, which was evaluated with data of two IoT use case.

Time: 29.01.2019, 11:00
Location: Room 103 in the Fraunhofer IGD, Fraunhoferstrasse 5, S3|05
Speaker: Moritz Fuchs (Advisor: Matthias Noll)
Title: "Registrierung eines Biopsieroboters zu bildbasierten Planungsdaten" (Bachelor thesis)

Time: 14.12.2018, 13:00
Location: Room 073 in the Fraunhofer IGD, Fraunhoferstrasse 5, S3|05
Speaker: Robert Königstein (Advisor: Arjan Kuijper)
Title: "Application of Voting Based Pose Estimation on 2D Contour Data" (Master thesis)
Abstract: Estimating 6DoF poses of objects is the fundamental skill, not only to understand the world around you, but to actually interfere and interact with it. This task is essential for a more flexible automation of any kind of pick and place tasks, especially in industrial environments. Nevertheless, in the early days of modern computer vision, the vast majority of research was aiming at the recognition or classification of objects. After the proposal of Ref. [DUNI10] and further work based on a Point Pair Feature approach [CTT+12, HLRK16, LH18], pose estimation then became feasible, fast and precise for 3D data.
Acquiring high quality 3D data is still expensive and a lot of computing power is needed for its processing. In comparison, 2D data (images) is easy to acquire and faster to handle. For industrial purposes, the pose estimation task reduces to the estimation of distinct object poses that are detected in a scene given some well-defined variations. For this, 2D data can be sufficient and hence, it is desirable to develop pose estimation algorithms that can work on plain image data. By the projection of the physical scene setup onto the camera sensor during the image acquisition process, valuable information is lost. A 6DoF pose contains information about the position and the rotation of an object in space. In 2D data, estimating position information is difficult, because distance and hence scale information is not available for an easy inference of the object position. For estimating the correct rotation of the object, restrictions apply in 2D, because visible angles and scale relations do not match the physical dimensions (shear is introduced by the camera projection). To overcome this lack of information and match 3D information of an object to 2D scene data, an object description with pair features, that describe local geometric object information effectively, can be considered.
The aim of this thesis is the estimation of 6DoF poses from 2D contour image data in an industrial context. Therefore, this work proposes a novel approach for 6DoF pose estimation on 2D data. A Point Pair Feature definition from Ref. [CTT+12] is adapted to two dimensional input data and the corresponding algorithm altered so that it can process 2D images. In this solution, a predefined object pose is detected and its pose is estimated from the contours of a scene image. For creating a discriminative object descriptor, the CAD model of the object is used. Local descriptions of scene point pairs are matched to a global object descriptor, each match voting for a 3DoF pose transformation. To get an estimation for the remaining 3DoF, depth and the camera position are discretised for the same process.
The results shown in this thesis demonstrate that PPFs can be adapted to be used in a pose estimation process on 2D data. Because of the discretisation of three of the six degrees of freedom, a higher precision of the pose estimations can be traded against computing time. With this approach, object detection and pose estimation for multiple object instances at a time is possible.

Time: 14.12.2018, 13:00
Location: Room 103 in the Fraunhofer IGD, Fraunhoferstrasse 5, S3|05
Speaker: Viviane Seidel (Advisor: Daniel Weber)
Title: "Simulation of 4D Programmable Textiles created by 3D Print on Prestressed Fabric" (Bachelor thesis)
Abstract: The objective of this thesis is to develop a simulation with the finite element method for three-dimensional imprinted, prestretched textile. For this purpose, methods necessary to model the intended behavior are conceived and tested. More specifically, methods are developed that should move the imprinted elements towards the stretched state, in which they have been when they got imprinted. A simple adjustment of the rest positions of the vertices as well as a modification of the derivative of the basis functions were found out not to be expedient. The application of an initial strain was discovered to be a functional strategy to maintain the stretched state. Adding a second layer of finite elements to the imprinted sections turned out to be a viable option to integrate the two different material characteristics of the components. However, this approach does not keep up the stretch of the imprinted areas, since the elements are still only connecting the same old vertices with the old initial rest states. Finally two options for the required perturbation of the simulation are introduced and their impact is analyzed. The combination of the respectively selected approaches results in a realistic simulation generating a visually convincing reproduction of the reality of 4D programmable textiles. Moreover, the linear model allows for an interactive modification of the resulting shape by modifying the scenario and adjusting material parameters.

Time: 10.12.2018, 15:30
Location: Room 220 in the Fraunhofer IGD, Fraunhoferstrasse 5, S3|05
Speaker: Namitha Chandrashekara (Advisor: Arjan Kuijper)
Title: "User-Centered Scientific Publication Research and Exploration in Digital Libraries" (Master thesis)
Abstract: Scientific research is the basis for innovations. Surveying the research papers is an essential step in the process of research. It is vital to elaborate the intended writing of state of the art. Due to the rapid growth in scientific and technical discoveries, there is an increasing availability of publications. The traditional method of publishing the research papers includes physical libraries and books. These become hard to document with the rise in the number of publications produced. Due to the above mentioned problem, online archives for scientific publications have become more prominent in the scientific community. The availability of the search engines and digital libraries help the researchers in identifying the scientific publications. However, they provide limited search capabilities and visual interface. Most of the search engines have a single field to search and provides basic filtering of the data. Therefore, even with popular search engines, it is hard for the user to survey the research papers as it limits the user to search based on simple keywords. The relationships across multiple fields of the publications are also not considered such as to find the related papers and papers based on the citations or references
The main aim of the thesis is to develop a visual access to the digital libraries based on the scientific research and exploration. It helps the user in writing scientific papers. A scientific research and exploration model is developed based on the previous information visualization model for visual trend analysis with digital libraries, and with consideration of the research process. The principles from Visual Seeking Mantra are incorporated to have an interactive user interface that enhances the user experience.
In the scope of this work, a research on Human Computer Interaction, particularly considering the aspects of user interface design are done. An overview of the scientific research, its types and various aspects of data analysis are researched. Different research models, existing approaches and tools that help the researchers in literature survey are also researched. The architecture and the implementation details of scientific research and exploration that provides visual access to digital libraries are presented.

Time: 12.11.2018, 16:00
Location: Room 073 in the Fraunhofer IGD, Fraunhoferstrasse 5, S3|05
Speaker: Ranveer Purey (Advisor: Arjan Kuijper)
Title: "Visual Trend Analysis on Digital Semantic Library Data for Innovation Management" (Master thesis)
Abstract: The amount of scientific data published online has been witnessing massive growth in the recent years. This has led to exponential growth in the amount of data stored in digital libraries (DLs) such as springer, eurographics, digital bibliography and library Project (dblp), etc. One of the major challenges is to prevent users from getting lost in irrelevant search results, when they try to retrieve information in order to get meaningful insights from these digital libraries. This problem is known as information overload. Other challenge is the quality of data in digital libraries. A quality of data can be affected by factors such as missing information, absence of links to external databases or data is not well structured, and the data is not semantically annotated. Apart from data quality, one more challenge is the fact that, there are tools available which help users in retrieving and visualizing the information from large data sets, but these tools lack one or the other basic requirements like data mining, visualizations, interaction techniques etc. These issues and challenges have led to increase in the research in the field of visual analytics, it is a combination of data processing, information visualization, and human computer interaction disciplines. The main goal of this thesis is to overcome the information overload problem and the challenges mentioned above. This can be achieved by using digital library named SciGraph by springer, which serves as a very rich source of semantically annotated data. The data from SciGraph can be used in combination with data integration, data mining and information visualization techniques in order to aid users in decision making process and perform visual trend analysis on digital semantic library data. This concept would be designed and developed as a part of innovation management process, which helps transforming innovative ideas into reality using a structured process. In this thesis, a conceptual model for performing visual trend analysis on digital semantic library data as part of innovation management process had been proposed and implemented. In order to create the conceptual model, several disciplines such as human computer interaction, trend detection methods, user centered design, user experience and innovation management have been researched upon. In addition, evaluation of various information visualization tools for digital libraries has been carried out in order to find out and address the challenges faced by these tools. The conceptual model proposed in this thesis, combines the usage of semantic data with information visualization process and also follows structured innovation management process, in order to ensure that the concept and implementation (proof of concept) are valid, usable and valuable to the user.

Guest lecture

Time: 06.11.2018, 11:00
Location: Room 073 in the Fraunhofer IGD, Fraunhoferstrasse 5, S3|05
Speaker: Prof. Mathieu Desbrun
Carl F Braun Professor of Computing and Mathematical Sciences
Title: "Subdivision Exterior Calculus for Geometry Processing"

Thesis presentations

Time: 30.10.2018, 14:00
Location: Room 220 in the Fraunhofer IGD, Fraunhoferstrasse 5, S3|05
Speaker: Daniel Hartung (Advisor: Hendrik Lücke-Tieke)
Title: "Interaktive Visualisierungen in flussdiagrammbasierten Programmierumgebungen" (Bachelor thesis)
Abstract: The analysis and processing of large amounts of data is a current challenge that domain experts as well as amateurs have to face. A possibility established for beginners and experts are flow based programming environments, which allow an user creating a pipeline for calculation. In existing solutions this pipeline is acyclical, which can limit the possibilities of interactive visualizations. In this thesis we want to create a prototypical concept which can weaken these limits. Thus it is possible to implement Brushing and Linking through the physical connection of visualizations and improve the possibilities of visual interaction. A qua- litative evaluation shows that these interactions can make a contribution to data analysis. Moreover this evaluation shows that cycles seems to be an intuitive solution for creating a Brushing and Linking setup. In this thesis we discuss existing solutions, the concept and the implementation of a prototype. Concluding we evaluate the prototype and point out some limitations.

Time: 26.09.2018, 13:00
Location: Room 103 in the Fraunhofer IGD, Fraunhoferstrasse 5, S3|05
Speaker: Florian Zouhar (Advisor: Ivo Senner)
Title: "Vector Based Web Visualization of Geospatial Big Data" (Bachelor thesis)
Abstract: Today, big data is one of the most challenging topics in computer science. To give customers, developers or domain experts an overview of their data, one needs to visualize these. They need to explore their data, using visualization technologies on high level but also in detail. As base technology, visualizations can be used to do more complex data analytic tasks. In case data contains geospatial information it becomes more difficult, because nearly every user has a well trained experience how to explore geographic information. These map applications provide an interface, in which users can zoom and pan over the whole world. This thesis focuses on evaluating one approach to visualize huge sets of geospatial data in modern web browsers. The contribution of this work is, to make it possible to render over one million polygons integrated in a modern web application which is done by using 2D Vector Tiles. Another major challenge is the web application, which provides interaction features like data-driven filtering and styling of vector data for intuitive data exploration. The important point is memory management in modern web browsers and its limitations.

Time: 14.09.2018, 15:00
Location: Room 324 in the Fraunhofer IGD, Fraunhoferstrasse 5, S3|05
Speaker: Maximilian Müller (Advisor: Jürgen Bernard)
Title: "Explorative Analyse von Multimodalen Patientendaten zur Erstellung und zum Visuellen Vergleich von Mehreren Patientenkohorten" (Master thesis)
Abstract: Maximilian Müller hat sich in seiner Masterarbeit 2 Kernproblemen bei der Analyse von Patientenhistorien gewidmet. Zum einen weist jede Patientenhistorie einen individuellen zeitlichen Verlauf auf, was die Selektion von Patientenkohorten erschwert. Zum anderen werden Patientenhistorien häufig durch eine ganze Reihe von Attributen charakterisiert, motiviert durch die Hoffnung der Ärzte auf neue, unerforschte Zusammenhänge in den Daten zu stoßen.
Der Ansatz von Maximilian Müller sieht drei Beiträge vor. Zum einen soll durch die gemeinsame visuell-interaktive Betrachtung von temporalen und multimodalen Eigenschaften von Patientendaten die Identifikation und Selektion Kohorten ermöglicht/vereinfacht werden. Weiter wird durch ein visuell-interaktives Interface der Vergleich temporaler und multimodaler Eigenschaften von Patientenkohorten möglich. Schließlich stellt Maximilian ein visuell-interaktives Werkzeug für die Erkennung und Analyse von Relationen zwischen temporalen und multimodalen Eigenschaften von Patientenhistorien vor.

Guest lectures

Time: 14.09.2018, 10:00
Location: Room 073 in the Fraunhofer IGD, Fraunhoferstrasse 5, S3|05
Speaker: Prof. Dr. Yasuyuki Matsushita      
Graduate School of Information Science and Technology, Osaka University
Title: "Semi-calibrated Photometric Stereo"
Abstract: 3D shape acquisition from images is of broad interest in computer vision and relevant industry applications. This talk introduces a "photometric" approach to high-fidelity 3D reconstruction, which enables shape recovery at the pixel-level detail. Specifically, this talk covers photometric stereo that takes as input images taken under different illumination conditions and estimates surface normal per-pixel. We introduce a new "semi-calibrated" photometric stereo, which is in-between calibrated and uncalibrated ones, and show that requirement about the light calibration for the conventional calibrated photometric stereo can be relaxed.
Bio: Yasuyuki Matsushita received his B.S., M.S. and Ph.D. degrees in EECS from the University of Tokyo in 1998, 2000, and 2003, respectively. From April 2003 to March 2015, he was with Visual Computing group at Microsoft Research Asia. In April 2015, he joined Osaka University as a professor. His research area includes computer vision, machine learning and optimization. He is on the editorial board of IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI), International Journal of Computer Vision (IJCV) and Encyclopedia of Computer Vision. He served/is serving as a Program Co-Chair of ACCV 2012, ICCV 2017, and a General Co-Chair for ACCV 2014 and ICCV 2021. He is a senior member of IEEE.

Speaker: Hiroaki Santo      
Graduate School of Information Science and Technology, Osaka University
Title: "Light Structure from Pin Motion"
Abstract: We present a practical method for the geometric calibration of point light sources which is based on a Lambertian plane and small needle pins at unknown positions that cast shadows on the plane. We show an interesting mathematical connection between pinhole cameras and point light sources which, combined with bundle adjustment, enables us to simultaneously estimate the point light position and the pin positions only from the observed shadow positions. This is directly analogous to traditional structure from motion where camera positions and 3D feature positions are estimated only from observed 2D feature positions
Our evaluation on simulated and real scenes shows that our method yields light estimates that are stable and more accurate than existing techniques while having a considerably simpler setup and requiring less manual labor.
Bio: Hiroaki Santo obtained his bachelor's and master's degree in computer science from Osaka University, Japan and is currently pursuing his PhD degree in Yasuyuki Matsushita's computer vision lab at Osaka University. His research interests are mainly in photometric stereo and machine learning.

Thesis presentations

Time: 05.09.2018, 16:00
Location: Room 324 in the Fraunhofer IGD, Fraunhoferstrasse 5, S3|05
Speaker: Isabelle Block (Advisor: Marcel Wunderlich, Tatiana von Landesberger)
Title: "Visuell-Interaktive Analyse von probabilistischer Krankheitsverbreitung mit Uneinheitlichen Übertragungswahrscheinlichkeiten" (Master thesis)

Time: 27.08.2018, 14:00
Location: Room 324 in the Fraunhofer IGD, Fraunhoferstrasse 5, S3|05
Speaker: Sanjaya Subedi (Advisor: Anirban Mukhopadhyay)
Title: "Generative Adversarial Networks for Image Synthesis and Video Prediction" (Master thesis)
Keywords: Deep Learning, Generative Adversarial Networks, Image Synthesis
Abstract: Generative adversarial networks (GANs) have gained a lot of research interest in recent years due to their ability to learn representations without having extensive annotated data. They capture statistical distribution of the training data, which can be used to synthesize samples from the learned distribution. They have been successfully used in synthesizing images that are almost indistinguishable from the real ones. Instead of manually writing equations for comparing similarity, or how an image matches a certain painting style, the network is able to automatically learn this from the data. This property is especially beneficial for image synthesis. In this thesis, we extend the idea of GAN for synthesizing Computed Tomography (CT) scan images from anatomical shapes and also synthesizing future frames of a video.
Physicians use the CT scan to diagnose, monitor and treat medical conditions. It is widely used for Radiotherapy Treatment Planning (RTP). Using X-ray measurements taken from multiple angles, cross-sectional images are produced to study the human body. CT scans typically involve high doses of radiation and repeated tests expose patients to significant level of radiation, increasing cancer risk. To enable CT-free RTP, synthesis of CT based on other imaging modalities is necessary.
Extending the idea of CycleGAN, we propose a model that can learn to synthesize CT from the shape of anatomy together with surrounding geometry. Our experiments on the dataset containing CT scans of femur and corresponding anatomical shapes from 90 patients show that the model is able to synthesize CT images accurately.
We also exploit Recurrent Convolutional Network in GAN framework to predict future frames in a video. Video prediction is a task where a plausible frame needs to be generated based on previous sequence of frames. It has its use in a wide range of domains including robotic planning in different kinds of surgical, industrial and commercial appliances, behavior analysis, and representation learning. Due to the wide range of activities that can be represented in a video, predicting video frames is a challenging problem that demands high-level understanding of scenes, physics, interactions and behaviors of different objects and many other aspects. In this thesis, we evaluate the proposed model on a surgical activity dataset for human motion modeling called JIGSAWS and on a synthetic dataset Moving MNIST. We also visualize salient features of the input frames to understand more about the network's prediction process.

Time: 27.08.2018, 10:00
Location: Room 073 in the Fraunhofer IGD, Fraunhoferstrasse 5, S3|05
Speaker: Simon Breitfelder (Advisor: Martin Ritz)
Title: "Example-based Synthesis of Seamless Texture Variations and Application to the Acquisition of Optical Material-Properties" (Master thesis)
Abstract: In dieser Arbeit wird ein Prozessfluss zur Erfassung, Synthese und dem Rendering von Approximate Bi-directional Texturing Functions (ABTF) erweitert, welche eine geringer-dimensionale Alternative zu Bi-directional Texturing Functions (BTF) darstellen. Im ersten Schritt werden hierbei Bildkorrekturen beschrieben und die Bildregistrierung des bestehenden Scanner-Aufbaus optimiert. Bei der Synthese von ABTF-Datensätzen ist es essentiell, auch die innere Struktur des Materials zu beachten, um für alle Beleuchtungsrichtungen eine konsistente Textur erzeugen zu können. Um große Flächen zu texturieren, wird zudem ein Verfahren vorgestellt, um nahtlose Texturvariationen zu erzeugen und diese beim Rendering zufällig auf der Oberfläche zu verteilen. Um den enormen Speicherbedarf von mehreren ABTF-Texturvariationen zu reduzieren, wird zudem ein Fitting auf ein Modell durchgeführt, welches die von ABTFs erfassten Oberflächeneigenschaften darstellen kann.

Internship presentation

Time: 10.08.2018, 10:30
Location: Room 074 in the Fraunhofer IGD, Fraunhoferstrasse 5, S3|05
Speaker: Timothy Dunn (DAAD RISE-scholarship holder and intern)
Clarkson University, Postdam, NY, USA
Title: Iimplementing a SDP (semidefinite programming) solver using a culip framework on CUDA" - a brief introduction into SDPs and explaining the crucial points of the implementation.

Thesis presentations

Time: 03.07.2018, 15:30
Location: Room 324 in the Fraunhofer IGD, Fraunhoferstrasse 5, S3|05
Speaker: Lisa Eisenhardt (Advisor: Kathrin Ballweg, Tatiana von Landesberger)
Title: "Exploration von anwendbaren Guidelines zur Visualisierung von Graphen für verschiedene Graphtypen" (Bachelor thesis)

Time: 29.06.2018, 11:00
Location: Room 074 in the Fraunhofer IGD, Fraunhoferstrasse 5, S3|05
Speaker: Tobias Dollenbacher (Advisor: Daniel Thuerck)
Title: "Patterns for the Distributed Solution of General MRF MAP problems by Dynamic Decomposition" (Bachelor thesis)
Abstract: A versatile tool for models in computer vision problems are Markov Random Fields. Since the solving of MAP problems on general pairwise Markov Random Fields is NP- hard, the big computational requirement for large Markov Random Fields make the solving process time consuming. The availability of high resolution images and cheap mass storage grow the input data for MRF MAP solvers. Hence, the need of an effi- cient solver that can deal with large general pairwise MRFs is stressed. In this thesis an approach to create such a solver by distributed > processing of huge MRF MAP problems using the mapMAP solver is described. Different patterns are introduced and used to dis- tribute the optimization data amongst the processes. The performance of the different strategies is evaluated and its advantages and weaknesses are shown. The underlying implementation uses MPI for the communication between the processes.

Time: 29.06.2018, 10:00
Location: Room 074 in the Fraunhofer IGD, Fraunhoferstrasse 5, S3|05
Speaker: Patrick Seeman (Advisor: Daniel Thuerck)
Title: "Soft Transparency for Point Cloud Rendering" (Master thesis)
Abstract: We propose a novel rendering framework for visualizing point data with complex structures or different quality of data. We characterize a point cloud using a per-point scalar field for differentiating parts of the dataset, i.e. based on the uncertainty of the points given by local normal variation or point density. Our rendering method uses the scalar field to render points as solid splats or semi-transparent spheres with non-uniform density to produce the final image. To that end, we derive a base model for integrating density in (potentially intersecting) spheres for both the uniform and non-uniform setting and introduce a simplified and fast approximation which yields interactive rendering speeds for millions of points. Having our method only rely on basic rasterization operations enables users to adjust rendering properties in real-time. The resulting interactive rendering of differently characterized point data leverages a clearer understanding of scenes in comparison with previous point splatting techniques and basic transparency rendering. Tests on several datasets with different characteristics and user studies substantiate our goal of easier scene understanding.

Time: 26.06.2018, 16:00
Location: Room 324 in the Fraunhofer IGD, Fraunhoferstrasse 5, S3|05
Speaker: Akshay Madhav Deshmukh (Advisor: Arjan Kuijper)
Title: "Automated User Evaluation Analysis for a Simplified and Continuous Software Development" (Master thesis)
Abstract: In today's world, computers are tightly coupled with the internet and play a vital role in the development of business and various aspects of human lives. Hence, developing a quality user-computer interface has become a major challenge. Well-designed programs that are easily usable by users are moulded through a regress development life cycle. To ensure a user friendly interface, the interface has to be well designed and need to support smart interaction features. User interface can become an Archilles heel in a developed system because of the simple design mistakes which causes critical interaction problems which eventually leads to massive loss of attractiveness in the system.
To overcome this problem, regular and consistent user evaluations have to be carried out to ensure the usability of the system.
The importance of an evaluation for the development of a system is well known. Most of the today's existing approaches necessitate the users to carry out an evaluation in a laboratory. Evaluators are compelled to dedicate the time in informing the participants about the evaluation process and providing a clear understanding of the questionnaires during the experiment. At the post experiment phase, evaluators have to invest a huge amount of time in generating a result report. On the whole, most of the today's existing evaluation approaches hogs up too much of time for most developments.
The main aim of this thesis is to develop an automated evaluation management and result analysis, based on a previous developed web-based evaluation system, which enables to elaborate the evaluation results and identify required changes on the developed system. The major idea is that an evaluation can be prepared once and repeated in regular time intervals with different user groups. The automated evaluation result analysis allows to easily check if the continued development lead to better results and if a bunch of given task could be better solved e.g. by added new functions or through enhanced presentation.
Within the scope of this work, Human-Computer Interaction (HCI) was researched, in particular towards User-Centered Design (UCD) and User Evaluation. Different approaches for an evaluation were researched in particular towards an evaluation through expert analysis and user participation. Existing evaluation strategies and solutions, inclined towards distributed evaluations in the form of practical as well as survey based evaluation methods were researched. A proof of concept of an automated evaluation result analysis that enables an easy detection of gaps and improvements in the system was implemented.
Finally, the results of the research project Smarter Privacy were compared with the manual performed evaluation.

Time: 22.06.2018, 10:00
Location: Room 103 in the Fraunhofer IGD, Fraunhoferstrasse 5, S3|05
Speaker: Johannes Heilmann (Advisor: Harald Wüst)
Title: "Visual-Inertial Model Target Tracking for Consumer Hardware" (Bachelor thesis)
Abstract: This thesis explores visual-inertial tracking for the application in Augmented Reality. Combining vision based tracking with data from inertial sensors like accelerometer and gyroscope can result in a faster and more robust tracking system. The two types of data complement each other well.
Tracking results in the form of 2D/3D correspondences from either a poster tracker or model tracker are combined with inertial data from the sensors of a Surface Pro 2 and fused in an Extended Kalman Filter. Different configurations are available. On the vision side there is a poster tracker using FAST and BRIEF for matching and the KLT for tracking. Alternatively a model tracker can be used. It builds a line model from a CAD model and tracks points on the lines in the image. On the inertial sensor side there are also two options. Either only gyroscope measurements can be used, or data from the gyroscope and an accelerometer can
be included.
The system is built to be used with consumer hardware. For this thesis a Microsoft Surface Pro 2 was used. This means that the data is not synchronised to a common clock and the inertial sensors are less accurate than those found in specialised hardware. The system is evaluated on a recorded image sequence and corresponding sensor data. The different parameters of the EKF models are tuned by experimentally minimising the RSME between EKF estimations and accurate baseline results.
It is shown that combining visual and inertial data allows the vision system to rely on tracking less features. This results in reduced computational cost. But the inertial sensors of the Surface Pro 2 are not accurate enough to allow a camera pose estimation based on inertial data alone in case the camera tracking fails.

Time: 15.06.2018, 16:00
Location: Room 324 in the Fraunhofer IGD, Fraunhoferstrasse 5, S3|05
Speaker: Philipp Schader (Advisor: Jürgen Bernard)
Title: "Visual-Interactive Optimization of 2D Data Layouts" (Visual Computing Lab)
Abstract: Die Darstellung von Datenobjekten in 2D mit Streudiagrammen (Scatterplots) eignet sich besonders gut um lokale und globale Strukturen (Outliers, Clusters vs. Korrelationen) in Daten zu erkennen.
Overplotting (zu viele Objekte an ähnlicher Position) gilt als eines der größten Probleme von Scatterplots
Der Ansatz von Philipp Schader zielt darauf ab Objekte über ein interaktives Masse-Feder Modell zu verschieben. Probleme bezüglich der Verfälschung der Darstellung hinsichtlich des ursprünglichen Daten-Arrangements löst Philipp über einen Visual Analytics Ansatz, bei dem auch Scagnostics Features zum Einsatz kommen werden. More details in the talk, including an interactive demo.

Honorable Mention

Honorable Mention for GRIS Paper at the "EuroVA Workshop for Visual Analytics", 04.06.2018, Brno, CZE

Dr. Jürgen Bernard and his collaborators received a honorable mention for their paper "Personalized Visual-Interactive Music Classification" at the EuroVis Workshop on Visual Analytics (Eurographics).
Together with Christian Ritter, Christian Altenhofen, Matthias Zeppelzauer, Arjan Kuijper, and Tobias Schreck, he proposed a visual-interactive and user-friendly interface that allows non-experts to train complex machine learning models (classifiers). Their use case is based on music tracks, reflecting one of the most popular types of personal data collections.

EuroVA Workshop

Thesis presentations

Time: 23.05.2018, 14:00
Location: Room 324 in the Fraunhofer IGD, Fraunhoferstrasse 5, S3|05
Speaker: Anand Vashishtha (Advisor: Anirban Mukhopadhyay)
Title: "Generalized U-Net for Multi-Channel Image Segmentation" (Master thesis)
Abstract: Image segmentation is a digital image processing technique which extracts important structural features of the image. Few of the practical applications of image segmentation are medical imaging, satellite imaging, remote sensing in the ecological domain and many others. U-Net is a robust encoder-decoder neural network for pixel-wise segmentation of grayscale biomedical images. However, U-Net does not generalize for multi-channel image segmentation; a necessity in multiple applications, for example, Wireless Capsule Endoscopy (WCE) in medicine, and habitat interpretation in ecology. In this thesis, we propose a deep learning method based on U-Net to segment multi-channel images in both ecology and medical
Biological diversity is decreasing at a rate of 100-1000 times pre-human rates, and to avoid species extinction; we need to understand factors influencing the occurrence of species. Fast, reliable computer-assisted tools can help to describe the habitat and thus to understand species habitat associations. This understanding is of utmost importance for more targeted species conservation efforts. Due to logistical challenges and time-consuming manual processing of field data, months up to years are often needed to progress from data collection to data interpretation. Image segmentation of vegetation dataset is one of the first tasks to interpret habitation preferences. Deep learning can be used to significantly shorten the time while keeping a similar level of accuracy. Here, we propose Habitat-Net: a novel Convolutional Neural Network (CNN) based method inspired from U-Net to segment multi-channel habitat images of rainforests. Compared to manual segmentation, Habitat-Net prediction is approximately 3K-150K times faster resulting in a significant increase of the processing time.
Similarly, a typical wireless capsule endoscopy procedure generates tens of thousands of images, resulting in a manual diagnosis of small bowel diseases laborious and time-consuming. Recent automatic approaches focus on learning based methodologies for patch-level localization of abnormalities, resulting in the argument about pixel-level localization marginal. This, however, results in a gross under-estimation as the distinctive bleeding pattern of many small bowel diseases are sprinkled over and covers only a small surface area. We present CE-Net a novel generalization of U-Net combining two regularization techniques, namely batch normalization and data augmentation for color image segmentation to automate pixel-level localization of the malign areas of WCE images. The advantages of CE-Net are demonstrated on a dataset containing 405 images across seven diseases, where CE-Net outperforms current state-of-the-art WCE image segmentation method in terms of Dice accuracy.

Time: 18.05.2018, 13:30
Location: Room 220 in the Fraunhofer IGD, Fraunhoferstrasse 5, S3|05
Speaker: Subhadeep Man (Advisor: Daniel Weber)
Title: "Jacobi Method for Distributed Sparse Linear Systems on Multiple GPUs using CUDA" (Maste thesis)
Abstract: Modern graphics processing units (GPUs) having many-core architecture are capable of accelerating simulation applications based on Parallel Differential Equation (PDE) tremendously. PDE solvers are the backbone of physics-based simulations and are used in wide variety of applications such as Computational Fluid Dynamics(CFD), computer games, augmented and virtual environments. Modern high-performance computing (HPC) architectures are equipped with multiple GPUs. This pushes the development of parallel algorithms towards solving fine resolution problems which are often inherently large for a single GPU. A Jacobi iterative solver is a popular PDE solver for Poisson equation. This thesis focuses on the scalability of a single GPU Jacobi solver over multiple GPUs. Transferring of data between multiple GPUs in a node remains the primary bottleneck and thus prevents from developing an efficient solver. Redesigning a solver which works on a single GPU and scaling it over multiple GPUs is challenging. Such a task would need identifying and exploiting parallelism opportunities and minimizing the communication latency when transferring data from one GPU to another. Existing methods used for multi-GPU communication are analyzed and optimal approaches are suggested for exploiting the inter-GPU memory bandwidth. Different domain decomposition methods are also suggested and the implications on performance are evaluated. Time taken to solve a problem increases with the resolution of the problem on a single GPU. Dividing the problem into smaller parts and distributing them to multiple GPUs could result in faster computation times. The implications of distributing such a problem are evaluated by measuring computation times and communication latency. Thereby, suggesting ways to optimize the solver by increasing the parallelism.

Time: 14.05.2018, 11:00
Location: Room 242 in the Fraunhofer IGD, Fraunhoferstrasse 5, S3|05
Speaker: Julian Puhl (Advisor: Matthias Noll)
Title: "Automatische Tumordetektion und Segmentierung im Bereich der Mundhöhle anhand der Symmetrie des menschlichen Kopfes" (Master thesis)
Abstract: Die Diagnose von Tumoren im Kopf-Hals-Bereich bedeuten für den Patienten in der Regel eine große Einschränkung der Lebensqualität. Aus diesem Grund wird versucht, die Behandlung sehr genau auf den jeweiligen Patienten abzustimmen. Ein wichtiges Hilfsmittel dabei sind Segmentierungen der Tumore
auf aufgenommenen Bilddaten. Für die Aufnahme kommt oft die Magnetresonanztomographie zum Ein-
satz. Für diesen speziellen Anwendungsfall existiert bisher kein automatisches Verfahren, welches gute
Ergebnisse liefert. An diesem Punkt setzt diese Arbeit an.
Es wird ein Verfahren vorgestellt, welches auf MRT-Aufnahmen automatisch Tumore erkennt und
segmentiert. Voraussetzung ist, dass die Aufnahme mit Verfahren zur Unterdrückung von Fettgewebe,
wie z.B. SPIR, erstellt wird und die Gewichtung der Aufnahme T1 ist. Außerdem muss dem Patienten
Kontrastmittel zur Hervorhebung des Tumorgewebes verabreicht worden sein. Das Verfahren nutzt die
Symmetrie des menschlichen Kopfes, um Asymmetrie und somit mögliche Tumore in der Aufnahme zu
finden. Hierfür wird die Symmetrieachse mittels Optimierungsverfahren und geeigneten Parametern be-
stimmt. Zusätzlich wird als weiteres Merkmal das durch Kontrastmittel in der Aufnahme aufgehellte
Gewebe verwendet. Diese beiden Merkmale kombiniert ergeben als Zwischenergebnis eine Maske, von
der die drei größten zusammenhängenden Komponenten extrahiert werden. Diese werden als Basis für
eine weitere Segmentierung verwendet. Hierbei kommt das Bereichswachstumsverfahren (engl. Regi-
on Growing) mit in dieser Arbeit entworfenem Kriterium zum Einsatz. Als Ergebnis erhält man bis zu
drei Segmentierungen möglicher Tumorkandidaten, von denen der Arzt entscheiden kann, welche er
verwenden möchte.
Evaluiert wurde das Verfahren auf 40 MRT-Aufnahmen, welche mit T1-Gewichtung, dem SPIR-
Verfahren zur Fettunterdrückung und verabreichtem Kontrastmittel aufgenommen wurden. Es standen
hierbei für alle Aufnahmen von Ärzten erstellte Segmentierungen der Tumore zur Verfügung. Als Ergeb-
nis wird ein Dice-Durchschnittswert von 0,64 und eine durchschnittliche Hausdorff-Distanz von 21,48
erzielt. Die Erkennungsrate liegt dabei bei sehr guten 90%.
Der Vorteil des Verfahrens ist, dass es vollautomatisch funktioniert und potentiell sogar mehrere Tu-
more finden könnte. Für die Endauswahl der Segmentierungen kann Expertenwissen einfließen und
bspw. können zusätzlich die gewählten Segmentierungen manuell weiter verfeinert werden. Dies bedeu-
tet im Normalfall eine große Zeitersparnis, da bei heutigem Arbeitsablauf die Segmentierungen oft noch vollständig per Hand erstellt werden.

Guest lecture (canceled!)

Time: 09.05.2018, 11:00
Location: Room 074 in the Fraunhofer IGD, Fraunhoferstrasse 5, S3|05
Speaker: Prof. Dr. Wolfgang Heidrich
King Abdullah University of Science and Technology (KAUST)
Title: "Space-Time Reconstructions in Fluid Imaging and Tomography"
Abstract: Computer vision are highly relevant for solving problems in scientific imaging, specifically fluid imaging and x-ray CT reconstruction. In the latter application, traditional tomographic imaging methods fall apart if the scan target undergoes deformation during the scan process. Such deformations may occur either either involuntarily (e..g in the form of heat expansion as the CT system heats up), or deliberately to analyze dynamic processes instead of static objects. Successful space-time reconstructions therefore require joint solutions of the tomography inverse problem and motion stabilization. Similar challenges occur in the context of fluid imaging, where the desired outcome is not a volume density, but a velocity vector field with physical constraints such as incompressibility.

In this talk I will highlight how both problems can be tackled with approaches developed in the computer vision community.

Wolfgang Heidrich joined King Abdullah University of Science and Technology in 2014 as the Director of the Visual Computing Center and a Professor of Computer Science. He is also a Professor (on leave) at the University of British Columbia. Dr. Heidrich received his PhD in from the University of Erlangen in 1999, and then worked as a Research Associate in the Computer Graphics Group of the Max-Planck-Institute for Computer Science in Saarbrucken, Germany, before joining UBC in 2000. Dr. Heidrich's research interests lie at the intersection of computer graphics, computer vision, imaging, and optics. His more recent interest is in computational imaging and display, focusing on hardware-software co-design of the next generation of imaging systems, with applications such as High-Dynamic Range imaging, compact computational cameras, hyperspectral cameras, to name just a few. Dr. Heidrich's work on High Dynamic Range Displays served as the basis for the technology behind Brightside Technologies, which was acquired by Dolby in 2007. Dr. Heidrich has served on numerous program

committees for top-tier conferences such as Siggraph, Siggraph Asia, Eurographics, EGSR, and in 2016 he is chairing the papers program for both Siggraph Asia and the International Conference of Computational Photography (ICCP). Dr. Heidrich is the recipient of a 2014 Humboldt Research Award.

Thesis presentations

Time: 06.04.2018, 11:00
Location: Room 074 in the Fraunhofer IGD, Fraunhoferstrasse 5, S3|05
Speaker: Fabian Czappa (Advisor: Stefan Gutthe)
Title: "Hardware zur Texturkompression" (Bachelor thesis)
Abstract: Aktuelle Computerspiele werden immer anspruchsvoller, die dargestellten Objekte lassen sich in Modelle und Texturen aufteilen, wobei die Texturen den gröÿeren Teil des Speicherplatzes verbrauchen. Der dedizierte Grafikspeicher allerdings ist im Normalfall fest auf der Grafikkarte verbaut und lässt sich nicht nachrüsten, was Entwicklungsstudios vor das Problem stellt, dass sie die Texturen nicht so hochauflösend verbreiten können, wie es möglich wäre. In dieser Bachelorarbeit präsentiere ich Erweiterungen für einen verlustfreien Kompressionsalgorithmus, sodass alle aktuell relevanten Texturen komprimiert werden können. Auÿerdem lassen sich durch diesen Ansatz zukünftig relevante Texturformate durch einfache Erweiterungen komprimieren. Der Algorithmus wird mit echten Daten getestet und diese Daten werden zur Analyse von realen Grafikkarten benutzt, um die Effizienz unter echten Bedingungen zu testen. Dabei verringert sich die Anzahl der nötigen Datentransfere zwischen dem Level 2 Cache und dem Videospeicher, wenn die Textur komprimiert wurde, um ungefähr 75%.

Time: 28.03.2018, 15:30
Location: Room 324 in the Fraunhofer IGD, Fraunhoferstrasse 5, S3|05
Speaker: Akshay Deshmukh (Advisor: Arjan Kuijper)
Title: "Automated User Evaluation Analysis for a Simplified and Continuous Software Development" (Master thesis)
Abstract: In today’s world, computers are tightly coupled with the internet and play a vital role in the development of business and various aspects of human lives. Hence, developing a quality user-computer interface has become a major challenge. Well-designed programs that are easily usable by users are molded through a regress development life cycle. To ensure a user friendly interface, the interface has to be well designed and need to support smart interaction features. User interface can become an Achilles heel in a developed system because of the simple design mistakes which causes critical interaction problems which eventually leads to massive loss of attractiveness in the system. To overcome this problem, regular and consistent user evaluations have to be carried out to ensure the usability of the system.
The importance of an evaluation for the development of a system is well known. Most of the today’s existing approaches necessitate the users to carry out an evaluation in a laboratory. Evaluators are compelled to dedicate the time in informing the participants about the evaluation process and providing a clear understanding of the questionnaires during the experiment. At the post experiment phase, evaluators have to invest a huge amount of time in generating a result report. On the whole, most of the today’s existing evaluation approaches hogs up too much of time for most developments.
The main aim of this thesis is to develop an automated evaluation management and result analysis, based on a previous developed web-based evaluation system, which enables to elaborate the evaluation results and identify required changes on the developed system. The major idea is that an evaluation can be prepared once and repeated in regular time intervals with different user groups. The automated evaluation result analysis allows to easily check if the continued development lead to better results and if a bunch of given task could be better solved e.g.
by added new functions or through enhanced presentation.

Within the scope of this work, Human-Computer Interaction (HCI) was researched, in particular towards User-Centered Design (UCD) and User Evaluation. Different approaches for an evaluation were researched in particular towards an evaluation through expert analysis and user participation. Existing evaluation strategies and solutions, inclined towards distributed evaluations in the form of practical as well as survey based evaluation methods were researched. A proof of concept of an automated evaluation result analysis that enables an easy detection of gaps and improvements in the system was implemented. Finally, the results of the research project Smarter Privacy were compared with the manual performed evaluation.


Time: 16.03.2018, 10:30
Location: Room 074 in the Fraunhofer IGD, Fraunhoferstrasse 5, S3|05
Speaker: Marc Streit
Johannes Kepler Universität (JKU) Linz
Title: "Leveraging Provenance for Storytelling, Reproducibility, and Recall"
Abstract:The primary goal of visual data exploration tools is to enable the discovery of new insights. To justify and reproduce insights, the discovery process needs to be documented and communicated. A common approach to documenting and presenting findings is to capture visualizations as images or videos. Images, however, are insufficient for telling the story of visual discovery, as they lack full provenance information and context. Videos are difficult to produce and edit, particularly due to the nonlinear nature of the exploration process.
Most importantly, however, neither approach provides the opportunity to return to any point in the exploration in order to review the state of the visualization in detail or to conduct additional analyses. In this talk, I will introduce our efforts to more tightly integrate biomedical data exploration with the presentation of discoveries. Based on provenance data captured during the exploration process, users can extract key steps, add annotations, and author 'Vistories', visual stories based on the history of the exploration. These Vistories can be shared for others to view, but also to retrace and extend the original analysis. I will demonstrate how such methods can increase the reproducibility of cancer research and drug discovery. The presented work is part of the Caleydo project (, which is a long-running collaboration between JKU Linz, Harvard University and the University of Utah.
Marc Streit is a tenured Associate Professor at Johannes Kepler University Linz in Austria where he leads the Visual Data Science group.
He finished his PhD at Graz University of Technology in early 2011 and moved to Linz later that year. In 2012 he was a visiting researcher at the Harvard Medical School. As part of a Fulbright scholarship for research and lecturing he was a visiting professor at the Harvard Paulson School in 2014. Marc also teaches courses at the Imperial College Business School and Salzburg University of Applied Sciences.
His scientific areas of interest include visualization, visual analytics, and biological data visualization, where he is particularly interested in the integrated analysis of large heterogeneous data.
Together with his team he develops novel visual analysis tools for cancer research, drug discovery, and other application domains. Parts of his research are embedded in the Caleydo open-source project (, where he is one of the project leaders and founding-members. He is a PI or Co-PI in multiple multiple research and industry projects and also a key researcher in the COMET K1-Centre Pro2Future. Since 2016 he is CEO and Co-Founder of the JKU spin-off company datavisyn.

Marc won Best Paper Awards at InfoVis'13, BioVis’12, InfoVis’11, GI’10 and Honorable Mention Awards at EuroVis'16, CHI'14, InfoVis'14 and EuroVis’12. He is a co-author of the Nature Methods Points of View column. In 2013 he co-edited the Special Issue on Visual Analytics in the IEEE Computer journal. Additionally, he actively contributes to the scientific community by serving on the organizing and program committee of several scientific events as well as by acting as a reviewer for high-quality journals and conferences. He was program chair of the IEEE Visualization in Data Science Symposium and papers and later general chair of BioVis, the Symposium on Biological Data Visualization.

Thesis presentations

Time: 14.03.2018, 14:00
Location: Room 072 in the Fraunhofer IGD, Fraunhoferstrasse 5, S3|05
Speaker: Clindo Devassy Kannanayikkal (Advisor: Alex Ulmer)
Title: "User-Centered Anomaly Detection in Network Data" (Masterthesis)
Abstract: Identifying anomalies in network traffic logs is a very challenging task for a network analyst. With the ever-increasing number of devices that can be connected to the network, the need for detecting
anomalies is at the peak. Usual techniques for detecting such anomalies include visual analysis of network data or applying automated algorithms. Both techniques have major drawbacks. Visual
analysis requires high expertise of the analyst, and automated detection algorithms produce high rates of false alarms.
In this work, both techniques are combined to improve the detection and reduce the workload of the analyst. The visual interface gives the network administrator the power to edit the predictions made by
the algorithms. The feedback from the network administrator are used by the algorithms to improve the performance of the detector and to reduce the false alarms. The system is tested and evaluated on a publicly available dataset which shows that the system achieves competitive performance.

Time: 07.02.2018, 16:30
Location: Room 048 in the Fraunhofer IGD, Fraunhoferstrasse 5, S3|05
Speaker: Yannic Fischler (Advisor: Martin Heß)
Title: "Physikalisch-basierte Animation deformierbarer Festkörper mit Eulerschem Ansatz" (Master thesis)
Abstract: Physikalische Systeme gehorchen partiellen Differentialgleichungen, die sich aus den Newtonschen Gesetzen ergeben. Um diese Differentialgleichungen numerisch lösen zu können müssen sie räumlich und zeitlich diskretisiert werden. Die Eulersche Diskretisierung, in der im Gegensatz zur Lagrangeschen Diskretisierung nicht der Körper sondern das Simulationsgebiet diskretisiert wird, wird seit einigen Jahren auch für die Simulation deformierbarer Festkörper eingesetzt. Diese Ansätze sehen sehr viel versprechend aus. In dem Vortrag wird ein Verfahren zur Simulation deformierbarer Festkörper mit Eulerschem Ansatz präsentiert und der Frage nachgegangen, wo das Verfahren Schwächen hat und wie diese gelöst werden können.

Time: 01.02.2018, 13:00
Location: Room 242 in the Fraunhofer IGD, Fraunhoferstrasse 5, S3|05
Speaker: Christian Gutjahr (Advisor: Silvia Rus)
Title: "Designing Self-Aware Textiles" (Master thesis)
Abstract: We are surrounded by textiles in our everyday live. Making them capable of local monitoring and computing is already a growing field of research in the area of Smart Home and Ambient Assisted Living. Equipping usual furniture with sensors and simple computational elements can provide useful information about the user and help with identifying emergency events, for instance fall recognition. This thesis investigates an approach to apply those ideas to textile materials worn by users by embedding inertial measurement unit sensors in a non-intrusive manner. In our approach, a simulation framework is used to ensure the highest possible accuracy while keeping the amount of sensors needed as low as possible. For this, a simulated sensor grid across the whole jacket was evaluated. Later, a prototype which uses the deformation of the jacket to provide valuable information about the current state of the jacket will be introduced. The presented use case to help find the jacket is just one idea on how to use the information gained by the sensor network.

Time: 26.01.2018, 14:00
Location: Room 242 in the Fraunhofer IGD, Fraunhoferstrasse 5, S3|05
Speaker: Matthias Mettel (Advisor: Andreas Braun)
Title: "Surface acoustic arrays to analyze human activities in smart environments" (Master thesis)
Abstract: This thesis introduces an approach for tracking three different activities, including their context extension, with a precision of 94% by using multiple pickup/piezo sensors. The mechanical waves, which are created by people touching various objects, can be recognized with these sensors. The combination of classical signal processing and current methods of machine learning enables the implementation of a processing pipeline for classifying these signal events and assigning the propagated event signal to its activity. Compared to the C4.5 CART and BayesNET classifiers,the best precision and performance balance is offered by the SVM classifier. The observed activities in this thesis are Walking, Closing a Cupboard and Falling. Especially Walking and Closing a Cupboard provide a good basis for extending the context. For a context expansion of Walking, the classification classes are split into the shoe types. Closing a Cupboard is divided into the cupboard instances, which have different positions and facing directions, in the environment. To avoid creating a Non class an Impact Filter is applied for preprocessing the recorded signals. The utilized main features are the RMS value and the Zero Crossings of the time-domain signal. They are extended by the FFT vector, statistical values like the mean and standard derivation of this vector as well as the index of the maximal FFT value. With this results, it is possible to lower the issues with common sensors like wearables and cameras. An additional advantage is that it can very easily be integrated in any environment.

Time: 11.01.2018, 15:30
Location: Room 220 in the Fraunhofer IGD, Fraunhoferstrasse 5, S3|05
Speaker: Joel Müller (Advisor: Christian Altenhofen)
Title: "Subdivision-Based CSM and its application to shape optimization" (Master thesis)
Abstract: While most common simulation methods rely on a discrete representation, modern modeling software typically employs a continuous representation. For the ability to perform a simulation, the design model needs to be converted by generating a discrete mesh, which is a time-consuming and error-prone process. In this thesis, this gap between CAD and CAE is bridged with a simulation method on the basis of a subdivision model, which is suitable for both modeling and simulation. This eliminates the need for mesh generation, significantly reducing the time required for each simulation. The underlying Catmull-Clark subdivision scheme allows for accessible modeling of complex three-dimensional shapes. Additionally, the high degree of continuity of the generated subdivision model allows for accurate approximations of functions with only a small number of degrees of freedom. The subdivision-based simulation method presented in this thesis is applied to solve a problem in the field of structural mechanics.
To assess the quality of the solutions generated with this method, it is compared to the Finite Element Method, which is a commonly used method in this area of application. The subdivision-based method is incorporated in a shape optimization workflow, in order to demonstrate its potential effect on the efficiency of integrated modeling and simulation.

Time: 18.12.2017, 16:30
Location: Room 048 in the Fraunhofer IGD, Fraunhoferstrasse 5, S3|05
Speaker: Masood Hussain (Advisor: Arjan Kuijper)
Title: "A Distributed Approach for Web-Evaluations of Desktop and Web-Based Applications" (Master thesis)
Abstract: The process of evaluation is an important concept, and it has a wide range of advantages. The purpose of evaluation can vary from project to project and usually it is used to check system usability, acceptability, and functionality. Evaluation can also be used to compare two existing systems, or to analyse the benefits of the new approach against the existing approaches. Today, there are many existing methods which can be used to perform evaluation. Among the existing methods, the most effective is the distance evaluation method because of being cost-effective. It is easy to manage the distance evaluation and to analyze the result because it follows the already defined procedure.
However, the distance evaluation method can only be used to evaluate web applications and is not applicable to desktop applications.
In this thesis, an approach is suggested that will enable the user to perform distance evaluation of the desktop applications. The approach extends the existing Web-based Evaluation of Information Visualization which is currently limited to web applications. The idea is to transfer the desktop application from the host application computer and display it in the web browser without compromising the security of the host computer. For the security purpose, the interaction of the participant is only limited to the application, and the participant cannot use any other feature of the remote host. The suggested approach successfully evaluate the desktop applications.

Time: 15.12.2017, 14:30
Location: Room 242 in the Fraunhofer IGD, Fraunhoferstrasse 5, S3|05
Speaker: Abdul Rehman Zafar (Advisor: Helmi Ben Hmida)
Title: "Enhancing User Experience in Internet of Things systems in terms of Smart Rule Management" (Master thesis)
Abstract: The number of devices in the Internet of Things systems is increasing day by day. With this increase in the number of devices, more and more data is being produced everyday. The quality of life of the end user is improving by using this data more efficiently. These data events are used to track any change in the state of smart devices. The applications of such systems include smart homes, health monitoring , traffic monitoring , energy consumption , logistics etc. More sophisticated systems are needed which can make sense of the data and perform right actions. Hence, a rule management layer is needed to effectively process these events. However, in the traditional systems, these programs are only written in the back-end by the developer of the application and are triggered automatically. So the end user of the application never has access or has very minimal interface to view and manage the rules. More human involvement contributes to success of an IoT system. Technically naive users usually view the whole IoT system in terms of user interface. Lacking the ability to visualize the rules greatly reduces the user acceptance of such systems and such systems are more likely to fail. That is why we should bring this rule management part to the front end of the system, rather than the back-end as it always has been in the traditional IoT systems. Thus, we need to increase the human acceptance of such systems by enhancing the user experience with a more user-friendly rule management system. Having such power for an ordinary end user is life changing. This thesis focus on enhanced the the user experience with respect to rules management in IoT systems. This has been done by reducing the complexity for technically naive end users and giving more flexibility to the technical users. This combination enhances the overall user experience and user acceptance with respect to rules management in IoT systems.

Time: 15.12.2017, 13:30
Location: Room 220 in the Fraunhofer IGD, Fraunhoferstrasse 5, S3|05
Speaker: Daniel Roth (Advisor: Daniel Weber)
Title: "Efficient Methods for Nonlinear Materials for Interactive Deformation Simulations based on Finite Elements" (Master thesis)

Time: 15.12.2017, 13:00
Location: Room 072 inthe Fraunhofer IGD, Fraunhoferstrasse 5, S3|05
Speaker: Ahmad Salahuddin (Advisor: Dirk Siegmund)
Title: "Vision Based Food Recognition" (Master thesis)

Time: 15.12.2017, 10:00
Location: Room 324 in the Fraunhofer IGD, Fraunhoferstrasse 5, S3|05
Speaker: Yeimy Paola Valencia Usme (Advisor: David Kügler)
Title: "Automatische Segmentierung der Paukenhöhle und des Antrum Mastoideum" (Master thesis)
Abstract: Die vorliegende Arbeit befasst sich mit der Segmentierung und dreidimensionalen  Rekonstruktion der wesentlichen Luftregionen des Felsenbeins. Für die Extraktion der Luftregionen ist nicht nur die Segmentierung , sondern auch die optimale Trennung der verbundenen Luftregionen erforderlich. Da nur wenige existierende Arbeiten den Fokus auf die Segmentierung der Lufteinschlüsse des Felsenbeins legen, sind die Untersuchungen dieser Thesis relevant. Es zeigt sich, dass die Segmentierung durch die hohe Variabilität der Anatomie der Luftregionen eine echte Herausforderung darstellt. Da die Implementierung von bildbasierten Algorithmen zu einer suboptimalen Segmentierung der inneren Lufteinschlüsse führen kann, wird eine modellbasierte Methode „Aktive Kontur“ in dieser Arbeit implementiert. Die Ergebnisse weisen darauf hin, dass die Qualität der Deformation für die dreidimensionale Rekonstruktion von Parameterangaben von der Aktiven Kontur abhängt. Es wird darüber hinaus gezeigt, dass eine existierende Methode beliebige Lufteinschlüsse im Felsenbein extrahieren kann, wenn ein 3D-Atlas mit segmentierten Luftregionen zur Verfügung steht. Diese Methode verwendet affine mit elastischen Registrierungsmethoden für die Segmentierung der Luftregionen. Um diese Verfahren zu untersuchen wurde ein Datensatz mit 22 CT-Daten als Atlas für jedes CT-Bild mit unterschiedlichen Luftregionen generiert. Hierfür wurde die Paukenkuppel, der Paukenkeller, der Paukenmittelraum, das Antrum Mastoideum, die Mastoidzellen manuell segmentiert. Es wurde festgestellt, dass die Visualisierung der Luftregionen im Felsenbein nicht nur zur präoperativen Planung der chirurgischen Eingriffe an der Otobasis geeignet ist, sondern dass sie sich auch zur Evaluierung der Fähigkeiten von Miniaturrobotern zur endoskopischen Untersuchung der Paukenhöhle eignen würde. Dies ist jedoch nicht Gegenstand dieser Arbeit.

Time: 15.12.2017, 09:00
Location: Room 324 in the Fraunhofer IGD, Fraunhoferstrasse 5, S3|05
Speaker: Martin Jastrzebski (Advisor: David Kügler)
Title: "Image-based Tracking of a surgical Microrobot" (Master thesis)
Abstract: The demand to Minimally Invasive Procedures has arisen in the last years due to its several
benefits when compared with traditional open procedure. Between the challenges of this kind of
procedures it is possible to point the exact tracking of the instruments as one very studied and with a great
impact on the success of these operations. In our work we present an ITK-based system that estimates the
pose of a surgical micro-robot based in images extracted during the procedure. The system uses
the 3D model of the instrument, the projections information and the volumetric data from the
patient to estimate not only the position and orientation of the instrument.
The correct behavior was verified during an evaluation of the main elements that affect the
registration accuracy and success. Although the large amount of images to be generated showed as
a bottleneck on the system performance, those tests showed promising results with around 0.35mm
of accuracy when using 2 or more images.

Time: 14.12.2017, 13:00
Location: Room 324 in the Fraunhofer IGD, Fraunhoferstrasse 5, S3|05
Speaker: Jan Fischer (Advisor: Marcel Wunderlich)
Title: "Visuell-interaktive Generierung von unsicheren Planungsdaten" (Bachelor thesis)

Time: 14.12.2017, 13:00
Location: Room 140 in the Fraunhofer IGD, Fraunhoferstrasse 5, S3|05
Speaker: Sufyaan Jakate (Advisor: Helmi Ben Hmida)
Title: "Enhancing User Experience with respect to Device Management in Internet of Things Domain" (Master thesis)
Abstract: There has been a steady increase in the use and development of the IoT devices. This is due to the increasing capabilities of the IoT devices and companies desire to provide better and easy services to the IoT users. An IoT system consists of multiple devices, which work, in cooperation to provide different services to the users. Since the numbers of IoT devices are increasing in the system, it is becoming difficult to manage all the IoT devices in the existing system. Consider the example of a Home IoT System, which consists of hundreds of devices, and sensors, which work in collaboration to provide the services. The management of large number of devices for a novice end user will be very challenging and complex task. The primary focus of this thesis is simplifying and reducing the complexity of managing the IoT devices with perspective of the novice end users, thus enhancing the overall experience for the end users.

Time: 11.12.2017, 10:00
Location: Room 324 in the Fraunhofer IGD, Fraunhoferstrasse 5, S3|05
Speaker: Christoph Bauer (Advisor: Johannes Fauser)
Title: "Segmentierung des Unterkiefergelenks und der Mastoidzellen in CT-Daten des Felsenbeins" (Bachelor thesis)
Abstract: Für eine gewebeschonende Operation an der Otobasis ist eine genaue Planung mithilfe von 3D Modellen vor dem Eingriff oft hilfreich. Zur Erstellung eines 3D Modells wird eine exakte Segmentierung der einzelnen Strukturen im Bereich der Otobasis benötigt. Diese einzelnen Strukturen werden aus medizinisch aufgenommenen Bildmaterial gewonnen. Für die Segmentierung des Felsenbeins und der zugehörigen Mastoidzellen wurde bereits ein Algorithmus entwickelt, der jedoch Probleme bei der Segmentierung der Mastoidzellen und der Entfernung des Unterkiefers besitzt. Diese Bachelorarbeit widmet sich diesen Problemen und segmentiert den Unterkiefer sowie die Mastoidzellen im Felsenbein. Zusätzlich wird ein Algorithmus zur Detektion des äußeren Gehörgangs beschrieben, der zu einer automatischen Initialisierung beiträgt. Als Eingabedaten erhalten wir ein dreidimensionales Grauwertbild der seitlichen Schädelbasis mit einer Angabe der zugehörigen Schädelhälfte und die Segmentierung der Risikostrukturen. Zur Lösung dieser Probleme wird eine Kombination von bildbasierten und formbasierten Segmentierungsalgorithmen genutzt.

Time: 05.12.2017, 16:30
Location: Room 073 in the Fraunhofer IGD, Fraunhoferstrasse 5, S3|05
Speaker: Vinod Singh Ramalwan (Advisor: Arjan Kuijper)
Title: "Visual Trend Analytics for Instant Analysis on Mobile Environments" (Master thesis)
Abstract: Big data counts as the oil of the 21st century. Massive collection and storage of data does not lead to new insights or knowledge. Hence, appropriate analysis methods and graphical tools are required to be able to extract a meaning from the data. In particular, the combination of data mining approaches together with visual analytics leads to real beneficial application to support decision making in business management. However, these big data analysis are predominantly performed with computers which are connected to high-resolution monitors. These high-res monitors are essential to show all available dimensions and facets of the data in multi-variate and analytical visualizations.
Perpendicular to the development of big data analysis, there is also a trend using mobile devices more often in daily business. Smartphones and tablets are more frequently used in mobile environments e.g. on business travels. Actually there exist rare solutions that enable a data analysis in principle on these devices; however, especially for trend analysis,
(currently) no mobile solutions exist. Due to the fact of small screens, imprecise interactions in visualizations via touch, slow internet connections and missing calculation performance on the devices leads to failure of classic analysis strategies.
There are currently three major challenges that limits practical use of visual trend analysis on these devices: (1) limitation of the result visualizations, due to approx. 5 inch displays on smartphones or approx. 9 inch displays on tablets, (2) associated with the small screen sizes there is even a much smaller interaction area with imprecise finger touch interaction, and (3) the reduced available performance, in particular due to 2-3 GB memory limitation. Due to these challenges most of the transformation of actual desktop solutions to mobile devices had failed some more references. It is essential to address the challenges and create new approaches that on the one hand takes in to account the limitations and on the other hand makes use of mobile device benefits, such as gesture interactions.
In this thesis, I present a scenario-based visual trend analysis which considers the raised issues and try to find novel solution such that the end user can comfortably use the visual trend analysis system on mobile environments. User evaluation is also conducted as part of the thesis so that one could find the difference between same system that runs on both desktop and mobile environment.

Time: 01.12.2017, 12:00
Location: Room 324 in the Fraunhofer IGD, Fraunhoferstrasse 5, S3|05
Speaker: Henry Krumb (Advisor: David Kügler)
Title: "Analyse des Trackingfehlers für Elektromagnetisches Tracking" (Bachelor thesis)
Abstract: In surgical procedures where there is no line of sight available, electromagnetic
tracking systems are useful to track the positions and orientations of
surgical tools in the inside of the patient's body.
Still the tracking accuracy is affected adversely by ferromagnetic objects
in the tracker's vicinity.
In this thesis, the effect of potential sources of error on small sensor
displacements was examined.
To enable further analysis, an etalon was developed using regular LEGO-bricks,
which do not interfere with the magnetic tracking.
Using the etalon and a defined measurement protocol, the sensor was placed
at several defined measurement points to measure the relative displacements
of multiple sensors that were compared with the ground truth provided by the
The assessments took place under laboratory conditions, in the environment of
a C-Arm and in an operating theater.
During the analysis of different measurement rates, an unexpectedly long
warm-up period was discovered.
The warm-up process causes a drift of position values by up to 0.1 mm.

Time: 09.11.2017, 15:00
Location: Room 072 in the Fraunhofer IGD, Fraunhoferstrasse 5, S3|05
Speaker: Andreas Rudolph (Advisor: Martin Hess)
Title: "Immersive Objektersetzung in augmentierter Realität" (Bachelor thesis)
Abstract: Der Einsatz von Virtual Reality (VR) und Augmented Reality (AR) zum Einblenden von Informationen ist auf dem Weg den Verbrauchermarkt zu erreichen. Ermöglicht wird dies durch neue Technologien und immer kompakter werdende VR/AR Geräte. Ein Einsatzszenario ist die virtuelle Objektersetzung von Einrichtungsgegenständen im Alltag. Wir stellen eine Studie vor, die die Microsoft HoloLens auf ihre Tauglichkeit auf diesem Gebiet untersucht. Dabei werden Einrichtungsgegenstände von ihren 3DRekonstruktionenn virtuell überlagert und deren Wirkung und Genauigkeit durch eine Benutzerstudie evaluiert. Von besonderem Interesse ist, wie erfolgreich Gegenstände virtuell verdeckt werden können und wie immersiv es auf den Benutzer wirkt. Wir zeigen, dass die Hololens eingeschränkt tauglich ist. Des Weiteren wird vorgeschlagen, wie einige Einschränkungen wie die manuelle Positionierung der 3D-Objekte und die begrenzte Immersion der Benutzer potentiell gelöst werden können.

Time: 03.11.2017, 14:00
Location: Room 242 in the Fraunhofer IGD, Fraunhoferstrasse 5, S3|05
Speaker: Mristi Rimal (Advisor: Helmi Ben Hmida)
Titlel: "Enhancing Visual Interface Aspects of Internet of Things Smart Object" (Master thesis)
Abstract: The Internet of Things, commonly known as "IoT", is the current ongoing evolution of the Internet. IoT encompasses all the services of a real-world object and therefore affects everyday human life. IoT development runs around finding ways to cooperate humans and objects. Humans are the producers and consumers of the IoT ecosystem. One key to the successful use of the IoT systems is more human involvement. Human users interact with the smart objects in the IoT ecosystem through the use of an interface. The interface of an IoT allows users to trigger actions in the smart objects, control the smart objects, and query the state of the smart objects. Technically naive users usually view the whole IoT system in terms of user interface. Any IoT development that does not address the user interface aspects is bound to fail. Recognizing the value of user interface, this thesis focuses on the enhancement of the visual interface aspects of IoT. This thesis provides an overview of the existing IoT platforms, their capabilities, characteristics, limitations and their comparison in terms of design requirements for a successful user interface. The limitations are then addressed and a user interface is proposed, combining all success factors which is evaluated with a user requirements survey and an implementation feedback questionnaire. The evaluation also follows possible future work.

Inaugural lecture

Time: 01.11.2017, 13:30
Location: Room 074 in the Fraunhofer IGD, Fraunhoferstrasse 5, S3|05
Speaker: Dr. Tatiana von Landesberger
Title: "Visualization for Data Communication: The Challenges in Creating a Picture Worth Thousand Words
Abstract: Data visualization communicates the data to the user for exploring unknown datasets, for confirming an assumed hypothesis about a dataset and for presenting results of an analysis. The data can stem directly from measurements, from simulations or can be a result of data modelling.
Nowadays, there are many tools and libraries to visualize data such as Tableau, Gephi, ESRI or D3. However, “simply” plotting the data on the screen may lead to several problems when reading and interpreting the data. Examples include data overplotting, “getting lost” in the data space during exploration, cognitive overburden and conveying data uncertainty.
The lecture will explain recent techniques for dealing with these visualization challenges. The techniques will be exemplified on data from various application domains such as transportation, journalism or biology.
Biography: Tatiana Landesberger von Antburg leads the Visual Search and Analysis group at Interactive Graphics Systems Group, TU Darmstadt. She finished her PhD in 2010 at this university. Her research interests cover visual analysis of spatio-temporal and network data from various application domains such as finance, transportation, journalism or biology.

Thesis presentations

Time: 25.10.2017, 11:00
Location: Room 324 in the Fraunhofer IGD, Fraunhoferstrasse 5, S3|05
Speaker: Markus Höhn (Advisor: Tatiana von Landesberger)
Title: "Design und Evaluation von Datenvisualisierungen mit mehreren Skalenebenen in Balkendiagrammen" (Bachelor thesis)
Abstract: Daten mit großem Wertebereich treten in der Praxis häufig auf. Beispiele hierfür sind Bruttoinlandsprodukte oder Einwohnerzahlen verschiedener Länder. In der Regel werden solche Daten mithilfe einer logarithmischen Skalierung dargestellt. Diese hat aufgrund ihrer fehlenden Linearität jedoch Nachteile bezüglich eines quantitativen Vergleichs zweier Werte.
In der vorliegenden Arbeit wird ein neuartiger Ansatz für Balkendiagramme aufgezeigt, welcher sowohl kleine, als auch große Werte angemessen darstellen kann. Um dies zu erreichen werden nicht nur die Höhe, sondern mit Breite und Hellwert auch weitere visuelle Variablen genutzt. Die dadurch entstehende Flexibilität in der Anpassung der Daten gewährleistet ein übersichtliches Abbilden großer Wertebereiche. Als weiteren Vorteil besitzt die vorliegende erarbeitete Visualisierung auch eine große Ausnutzung der vorhandenen Fläche.
Um die Nutzbarkeit des vorliegenden Entwurfs zu evaluieren, wird eine Benutzerstudie durchgeführt, welche die Vorteile hinsichtlich quantitativer Vergleichbarkeit zweier Werte bestätigt und das neuartige Width-Scale-Balkendiagramm als gute Alternative zu üblichen Balkendiagrammdarstellungen bestätigt.

Time: 25.10.2017, 10:45 / 11:15
Location: Room 048 in the Fraunhofer IGD, Fraunhoferstrasse 5, S3|05
Speaker: Daniel Ceballos / Roman Rempel (Advisor: Johannes Fauser)
Title: "Algorithm Configurator for Medical Image Processing & Interactive Motion Planning" (Praktikum Visual Computing)
Abstract: Minimal-invasive Chirurgie an der Otobasis erfordert einen präzisen präoperativen Planungsprozess. Hierbei werden für die Operation relevante Organe zunächst in CT-Daten segmentiert, um ein 3D Modell der Landmarken und Risikostrukturen zu erhalten. In dieser Umgebung werden anschließend Zugangswege für chirurgische Instrumente geplant, die dann von einem Bohrroboter aus dem Knochen gefräst werden. Um den präoperativen Planungsprozess in einer graphischen Benutzeroberfläche möglichst einfach und intuitiv zu ermöglichen, erarbeiteten die beiden nachfolgenden Praktika Erweiterungen zu dieser GUI im Bereich Segmentierung und Pfadplanung:

Im Projekt "Algorithmenkonfigurator" wurde damit begonnen, einen Wrapper für eine allgemeine Image Processing Pipeline und eine interaktive Visualisierung dieser Pipeline als Graph zu implementieren (vgl. MeVisLab, XPIWIT), um die für die Segmentierung der Risikostrukturen benötigten Algorithmen (z.B. aus pdmLib, ITK, etc.) effektiv weiterzuentwickeln.
Im Projekt "Interactive Motion Planning" wurden Interaktionsmöglichkeiten (ImplicitPlaneWidget, etc.) aus VTK auf die Bedürfnisse von Motion Planning Problemen zugeschnitten, um Start- und Zielregionen von Pfadplanungsalgorithmen interaktiv in der 3D Visualisierung der Anatomie zu platzieren.

Time: 25.10.2017, 10:00
Location: Room 324 in the Fraunhofer IGD, Fraunhoferstrasse 5, S3|05
Speaker: Johann Reinhard (Advisor: Alan Brunton)
Title: "Discrete Medial Axis Transform and Applications for 3D Printing" (Bachelor thesis)
Abstract: 3D printing is becoming a more commonly used manufacturing process, both for industrial and consumer use, with ever increasing capabilities and areas of application. These opportunities also introduce higher expectations on the quality of the resulting prints, generally in terms of the resulting shape and appearance of the object, but also rigidness and structural integrity. Detecting characteristics in a model that are a source of errors opens up possible approaches to mitigate or eliminate these errors before printing it. One such characteristic are thin structures that can lead to missing or deformed shapes, changes in the appearance of full color prints or fragile structures that break during post-processing steps.
The aim of this work is to detect thin structures using the discrete medial axis, representing the centers of a shape. In order to compute the discrete medial axis a discrete medial axis transform based on image processing techniques is implemented in the Cuttlefish 3D printer driver. The result for different models are assessed and possible correlations of the medial axis and thin structures evaluated. Possible applications of the medial axis or filtered medial axis are proposed and discussed.

Time: 25.10.2017, 10:00
Location: Room 048 in the Fraunhofer IGD, Fraunhoferstrasse 5, S3|05
Speaker: Heiko Reinemuth und Hendrik Pfeifer (Advisor: Jürgen Bernard)
Title: "Visual-Interactive Preprocessing of Multivariate Time Series" (Praktikum Visual Computing)
Abstract: Datenqualität ist ein zentrales Problem für die meisten Ansätze zur Datenanalyse. Heiko Reinemuth und Hendrik Pfeifer haben sich dem speziellen Problem des Preprocessings von multivariaten Zeitserien angenommen. In ihrem Praktikum erarbeiteten sie visuell-interaktive Lösungen für das Preprocessing von multivariaten Zeitserien. Ihre Lösungen zeichnen sich durch eine besondere Praxisnahe aus: so arbeiteten Heiko und Hendrik insgesamt mit Datensätzen aus drei unterschiedlichen Anwendungsgebieten: menschliche Bewegungsdaten, Wetterdaten, sowie EEG-Daten.

Time: 25.10.2017, 09:30
Location: Room 048 in the Fraunhofer IGD, Fraunhoferstrasse 5, S3|05
Speaker: Maximilian Müller (Advisor: Jürgen Bernard)
Title: "Visualization of Time-Oriented Information for Large Patient Databases" (Praktikum Visual Computing)
Abstract: Patientendaten werden in Zukunft Forschung, sowie Therapierung im medizinischen Kontext revolutionieren - da sind sich viele Experten einig. Die Visualisierung von vielen Patientendaten (auf wenig Display Space), könnte diese Entwicklung unterstützen, stellt sich jedoch als schwierig heraus. Patientendaten besitzen temporale, sowie multivariate Information - zudem ist jeder Patient einzigartig - kein einfaches Setting für die Visualisierung vieler Patienten. Basierend auf existierender Related Work im Bereich Prostatakrebs, hat sich Maximilian Müller auf die Visuelle Analyse von temporalen Aspekten in Patientenhistorien konzentriert. Er illustriert vier verbleibende Probleme der Related Work - und stellt entsprechend seine Lösungsvorschläge vor.

Tableau Workshop

Time: 24.10.2017, 15:20 - 17:20
Location: Room 074 in the Fraunhofer IGD, Fraunhoferstrasse 5, S3|05
Speaker: Christian Bersch (Tableau Software)
- The Beautiful Science of Data Visualisation
- Real-life case study
- Hands-on training on Tableau Desktop
More information:

Thesis presentations

Time: 24.10.2017, 14:00
Location: Room 072 in the Fraunhofer IGD, Fraunhoferstrasse 5, S3|05
Speaker: Alexandros Roussos (Advisor: Martin Hess)
Title: "Massiv Parallele Lösung des Transportproblems" (Bachelor thesis)
Abstract: Das lineare Zuordnungsproblem und allgemeiner das Transportproblem können als fundamentale Werkzeuge im Bereich Computer Vision und mathematische Bildverarbeitung gesehen werden. Für glatte kostenfunktionen in R^n, im speziellen für den quadratischen euklidischen Abstand, existieren spezielle Lösungsverfahren. Dies ist jedoch nur eine sehr eingeschränkte Klasse von Problemen und die verwendeten Differentialgleichungen und partiellen Differenzialgleichungen sind zum Teil numerisch nicht stabil. Für das lineare Zuordnungsproblem existieren zwei klassische Algorithmen: die ungarische Methode und der Auction Algorithmus, wobei der Zweite parallelisiert und auf das allgemeine Transportproblem erweitert werden kann. Um den Auction Algorithmus weiter zu beschleunigen, muss man die möglichen Transportpfade effzient einschränken, ohne die optimale Lösung zu verpassen. Zwei Verfahren hierzu sind ein Multiscalenansatz und der sogenannte Shortcuts Algorithmus, bei dem Ideen aus der ungarischen Methode in den Auction Algorithmus integriert werden. Basierend auf den Multiscalen Ansätzen von Bernhard Schmitzer wurde ein massiv paralleler Lösungsalgorithmus für das allgemeine Zuordnungsproblem entwickelt. Das Verfahren wurde mit Problemen unterschiedlicher Größe in R^n getestet und mit einer Referenzimplementation verglichen. Speziell geht es hier also um die Frage, welches Verfahren effzient parallelisiert werden kann und welche Probleme dabei gelöst werden müssen.

Time: 23.10.2017, 09:00
Location: Room 324 in the Fraunhofer IGD, Fraunhoferstrasse 5, S3|05
Speaker: Moritz Weissenberger (Advisor: Johannes Fauser)
Titel: "Interaktive Analyse von geplanten medizinischen Bohrkanälen" (Bachelor thesis)
Abstract: Eine offene Operation an Mittel- oder Innenohr ist für den Patienten immer mit Nachteilen verbunden, so können zum Beispiel postoperative Schmerzen oder Entzündungen den Krankenhausaufenthalt verlängern. Auch gehen offene Operationen mit Narben einher, die zu Komplikationen führen können. Naheliegend ist da der Wunsch, mehr in Richtung minimal-invasiver Verfahren zu forschen. Ein Anwendungsbeispiel für einen solchen minimalinvasiven Eingriff ist der Einsatz eines Cochleaimplantates (Hörprothese für Gehörlose mit Innenohrschädigung) in das Felsenbein. Doch minimal-invasive Verfahren benötigen ein hohes Maß an Vorausplanung. Ein Teil dieser Planung beinhaltet unter anderem die Auswahl des optimalen Bohrpfades. Da man aber nicht anhand nur eines Kriteriums den perfekten Pfad für jeden Patienten ausmachen kann, bedarf es einer Gewichtungsfunktion, die mehrere Variablen einbezieht, und die vom Arzt auf jeden Patienten passend zugeschnitten werden kann.Weiterhin wird eine Visualisierung jedes Punktes des Pfades für den Arzt benötigt, die es ihm ermöglicht, die Strukturen entlang des Pfades detailliert zu kontrollieren. Diese Arbeit befasst sich mit verschiedenen Kriterien mit deren Hilfe ein Pfad bewertet werden kann, wie diese visualisiert werden können, und wie ein Einblick in jeden Abschnitt des Pfades gewährt werden kann. Im Detail sind diese Kriterien der Abstand zu den Risikostrukturen, die Geradheit des Pfades, der Ankunftswinkel am Zielpunkt und die Struktur des Knochens auf dem Weg. Das wichtigste dieser Kriterien ist dabei wohl der größte minimale Abstand entlang des Pfades, wobei zwischen den einzelnen Strukturen eine zusätzliche Gewichtung sinnvoll erscheint. Als zweites wird die Geradheit des Pfades untersucht, da der Roboter nur bohrt und es bei einem geraden Pfad leichter für den Arzt ist, die anschließende Operation vorzunehmen. Aus einem ähnlichen Grund ist der Winkel beim Ankommen am Zielpunkt wichtig, denn wenn der Roboter zu stark von dem geplanten Ankunftswinkel abweicht, die sich anschließende OP für den Arzt verkompliziert wird. Anschließend wird die Dichte und Struktur des Gewebes entlang des Bohrpfades als möglicher Gewichtungsfaktor diskutiert und mögliche Probleme werden erläutert, die sich aus entlang des Pfades liegenden Luftlöchern ergeben. Gleichzeitig wird über möglichen Probleme, die von Luftlöchern entlang des Pfades ausgehen, berichtet. Weiterhin werden alle Ergebnisse der vorherigen Abschnitte zusammengefasst und der optimale Pfad anhand der gewählten Gewichte ermittelt. Darüber hinaus werden Möglichkeiten erörtert, anschaulich Einstellungen an den Gewichten vorzunehmen. Abschließend wird zur Visualisierung der Pfade für den Arzt eine Möglichkeit aufgezeigt mit der durch die CTBilder, orthogonal zum Pfad, gescrollet werden kann. Dabei werden diese CT Bilder in die 3D-Darstellung der Risikostrukturen eingeblendet werden.

Time: 17.10.2017, 10:30
Location: Room 242 in the Fraunhofer IGD, Fraunhoferstrasse 5, S3|05
Speaker: Lisa Scherfl (Advisor: Florian Kirchbuchner)
Title: "Human Behavior Analysis and Predicition Based on a Smart Floor" (Bachelor thesis)
Abstract: Older adults have the desire to live independently in their own homes for as long as possible. The development of sensor technologies in Smart Homes support this aim by providing sufficient security standards in case of emergencies. For example, a call of emergency can be triggered if a fall of a person is detected by sensors hidden in the floor. However, it is often not only about urgent situations, but also about gradual changes in behavior. Especially when a user is not able to follow his or her daily routine, long-term activity recognition based on location tracking allows for early detection of diseases such as Alzheimer's and dementia and can generally reveal a decrease in the ability to live independently. The focus of this work was the investigation of health related activities and their most accurate measurement only using an intelligent floor based system. Based on these considerations, a method to extrapolate from the collected sensor data to the chosen values is proposed. In addition, a model to detect gradual changes in these health indicators is developed and tested on the smart floor in the Living Lab of Fraunhofer IGD as well as in two apartments in everyday life. The findings of these thesis show a way of using smart floors for health monitoring. The applicability in everyday life could not be shown due to independent problems with the location tracking of the floor during the evaluation period and the lack of additional data for the validation. However, the evaluation under testing conditions showed promising results and an untapped potential of smart floors in health monitoring.

Time: 17.10.2017, 09:30
Location: Room 242 in the Fraunhofer IGD, Fraunhoferstrasse 5, S3|05
Speaker: Tabea Hartl (Advisor: Florian Jung)
Title: "Ein interaktives Segmentierungsverfahren für Strukturen in medizinischen Bilddaten unter Verwendung eines Random Walk Ansatz" (Master thesis)
Abstract: Head and neck tumors are a dangerous and invasive disease, that is why it is eminent to proceed a detailed diagnosis and treatment planning of the patient. An important part of this diagnosis and treatment planning is the segmentation of the tumor. For that it is common to use magnetic resonance imaging data. Currently and in this special case there are no good solution bringing procedures.
This thesis introduces an interactive segmentation algorithm for tumors in the head and neck region, which uses a random walk approach. The algorithm applies initial markers to create a segmentation of the tumors even though adjacent structures have similar intensity values. For reasons of efficiency the similarity to a problem of the circuit theory is used. Therefore, the image data is transformed in a graph, which is made of nodes and edges. Based on this graph a linear system of equations is created to compute the solution of the random walk approach.
The algorithmdelivers good results for tumors in the head and neck region in MR data. We tested the algorithm for 18 different tumors with two different sets of initial markers for each tumor. The median of the Hausdorff distance from all segmentations is 8.83 and the Dice coefficient is 0.71. Further, the approach was evaluated in CT data from four different structures and for lymph nodes in MR data.
A big advantage of this procedure is the possibility to use expert knowledge to improve the results. Most other procedures do not allow this possibility. The algorithm provides useful segmentations of tumors and lymph nodes in MR data. Furthermore, the algorithm is generic and can also be used for other structures and modalities.

Time: 16.10.2017, 15:15
Location: Room 048 in the Fraunhofer IGD, Fraunhoferstrasse 5, S3|05
Speaker: David Sessler (Advisor: Jürgen Bernard)
Title: "Visual-Interactive Learning of Time Series Similarity" (Master thesis)
Abstract: Similarity Functions are essential for many unsupervised algorithms used in Data Mining, Machine Learning, and Information Retrieval. The identification of meaningful similarity functions for particular data types, application goals, and information-seeking behaviors is challenging and tedious for data scientists, and particularly for domain experts without expertise in data science. In many cases similarity functions first need to (pre-) process and transform data into appropriate features spaces, we refer this process to as a workflow. David Sessler takes similarity search for time series to a new direction with the presentation of a visual-interactive system that automatically learns similarity functions for time series. The input for the learning process is directly expressed by users in a visual-interactive way: by assigning similarity scores for pairs of time series (labeling). According to the similarity labels, the system automatically assesses the performance of workflows for time series processing and similarity calculation. Visual-interactive views support users in the identification of meaningful workflows (and in particular: processing steps included in the workflows). A visual-interactive retrieval component allows the validation of identified similarity functions, as well as the iterative refinement (feedback loop). With the system, domain experts now have a means to create meaningful similarity functions for time series within minutes. The system shifts similarity search for time series from a tedious offline process to an iterative online process even for non-experts.

Time: 16.10.2017, 14:30
Location: Room 048 in the Fraunhofer IGD, Fraunhoferstrasse 5, S3|05
Speaker: Markus Lehmann and Martin Müller (Advisor: Jürgen Bernard)
Title: "User-based Strategies for Selecting Objects for Labeling" (Praktikum Visual Computing)
Abstract: Labeling objects is an important task for Machine Learning and for Visual Analytics as well. In a previous work, we showed that the user-based selection of objects for labeling data can compete with the performance of model-based selection strategies such as Active Learning strategies (well-known in the Machine Learning community). In this practical course, Markus and Martin implemented ten observed strategies of users for the selection of objects for labeling, when making use of visual-interactive interfaces for data exploration. Markus and Martin will assess the performance of a theoretical upper limit of performance (knowing the ground truth in advance) and compare it with performances of the ten user-based selection strategies. Finally, the students will propose ensemble-based strategies which combine the strengths of the ten user-based strategies. The driving question: will ensemble-based strategies further increase the performance of user-based strategies for the selection of objects for labeling?

GRIS Workshop

GRIS Workshop at the VIS Conference: “Visual Analytics in Healthcare” (VAHC)

Date: 2017, October October 1st / 2nd
Location: Phoenix, Arizona
Affiliated with IEEE VIS 2017
German abstract:
Dr. Jürgen Bernard ist einer der Organisatoren des diesjährigen VAHC Workshops auf dem Experten aus der Visual Analytics Forschung und der Medizin zusammentreffen werden. Gemeinsam wird nach visuell-interaktiven Lösungen zur Analyse von Daten aus dem medizinischen Kontext gesucht, um Problemen dieser speziellen Anwendungsdomäne entgegenzutreten. Neben speziellen Herausforderungen denen die Mediziner gegenüberstehen, lassen sich im Healthcaresektor nicht selten klassische „Big Data“ Problematiken wiederfinden: sehr große, heterogene Datenmengen, die sich zudem häufig verändern – wie zum Beispiel die Historien von Patienten. Ein zentrales Analyseziel ist daher die Unterstützung von Forschung und Praxis in der medizinischen Behandlung.


Thesis presentations

Speaker: Johannes Merz (Advisor: Roman Getto)
Title: "Simplified definition of parameter spaces of a GML model using sketch-based interaction" (Master thesis)
Abstract: This Master Thesis presents a novel technique to intuitively insert meta-parameters into a procedural GML model with the help of sketch-based interaction. A GML model consist of a sequence of procedural modeling commands, for example extrusions. These are called with a set of local offset values, which can be converted to global space and anchored in the surface mesh of the model. As the global positions do not necessarily lie on the mesh surface, this is done by finding reference vertices. The system uses a mesh deformation technique to deform the evaluated surface of the model and creates a progression of intermediate target meshes. During the deformation, the reference vertices provide the global offset positions, whose path can be approximated by a B-spline. By exchanging the initial values of the commands by this B-spline, a continuous parameter space of the meta-parameter is defined. The deformation process is supported by a mesh segmentation to create pre-defined deformation targets for the user. Using intuitive sketch-based methods, these can be easily adapted to the users needs. The results show that the system closely imitates the mesh deformation with the help of the modeling commands. Furthermore, the system was evaluated to be intuitive and easy to use.

Guest lecture

Time: 18.08.2017, 13:00
Location: Room 074 in the Fraunhofer IGD, Fraunhoferstrasse 5, S3|05
Speaker: Prof. Dr. Daniel Scharstein
Charles A. Dana Professor of Computer Science
Middlebury College
Title: "New Ideas for Stereo Matching of Untextured Scenes"
Abstract: Two talks for the price of one! I will present two radically different approaches to the challenging problem of stereo matching of scenes with little or no surface texture.
First, I will discuss how surface orientation priors can be added to the popular semi-global matching (SGM) algorithm, which significantly reduces errors on slanted weakly-textured surfaces. The orientation priors serve as a soft constraint during matching and can be derived in a variety of ways, including from low-resolution matching results and from monocular analysis and Manhattan-world assumptions.
Second, we will examine the pathological case of Mondrian Stereo – synthetic scenes consisting solely of solid-colored planar regions, resembling paintings by Piet Mondrian. I will discuss assumptions that allow disambiguating such scenes, present a novel stereo algorithm employing symbolic reasoning about matched edge segments, and discuss how similar ideas could be utilized in robust real-world stereo algorithms for untextured environments.

Thesis presentations

Time: 16.08.2017, 10:00
Location: Room 324 in the Fraunhofer IGD, Fraunhoferstrasse 5, S3|05
Speaker: Andrei Stefanov (Advisor: David Kügler)
Title: "Instrument Tracking with Convolutional Neural Networks" (Master thesis)
Abstract: The department GRIS is part of the joint MUKNO project, which investigates minimally invasive surgical procedures at the otobasis. The main idea of this project is to use a tiny robot to drill canals along a planned path. For the navigation during the operation it is essential to track the current position and orientation of the robot at any given time. For this scope a hybrid system should be used which consists of an electromagnetic tracking system and a c-arm X-ray unit. The task of this master thesis is to develop an automatic procedure to determine the position and orientation of the robot in intraoperatively acquired X-ray images using Convolutional Neural Networks. In order to get a better understanding of the concept of CNNs the task at hand is at first greatly simplified, after which the complexity of the problem is increased step by step. The final iterative solution described in this thesis achieves sub-pixel and sub-degree accuracy for the prediction of the position and orientation with a decent runtime of less than 2 seconds. Both the training of the CNNs and the validation of the implemented method were conducted on Digitally Reconstructed Radiographs, which allow the generation of huge amounts of image data with automatic annotation of the ground truth.

Time: 11.08.2017, 11:00
Location: Room 242 in the Fraunhofer IGD, Fraunhoferstrasse 5, S3|05
Speaker: Dhanashree Joshi (Advisor: Silvia Rus)
Title: "Emotion Detection By Evaluating Activities For Smart Home Appliances" (Master thesis)

Time: 09.08.2017, 16:00
Location: Room 073 in the Fraunhofer IGD, Fraunhoferstrasse 5, S3|05
Speaker: Dimitar Tomov Dimitrov (Advisor:Arjan Kuijper)
Title: "Non-Local Bayes Denoising of Digital Images on GPU using OpenCL" (Master thesis)
Abstract: Image sensors or lens inevitably produce a wide range of degradations in digital images. Among these degradations are various types of noise. Image noise is undesirable, random information added to the image that can cause significant quality degradation. Therefore, image denoising methods are very often the first step of an image processing chain.
Non-Local Bayes (NL-Bayes) is a state of the art image denoising algorithm that provides effective noise reduction with good preservation of image details. Although efficient CPU implementations of this method exist, they still have performance issues. Nowadays, GPUs become a general-purpose computational devices and provide their power for more generalized tasks. Their massively parallel capabilities can significantly improve the performance of computationally expensive programs. However, due to the specific execution and memory model, algorithms for GPUs require special design that divides the computational task into as many independent pieces as possible.
This thesis proposes solutions to perform NL-Bayes image denoising on the GPU using the OpenCL standard. Our study evaluates quantitatively and qualitatively the proposed implementations by making comparisons with an existing CPU implementation. We show that a GPU-based solution is six times faster compared to the CPU-based solution when maximal denoising quality is demanded. Additionally, we describe general problems in the parallelization of such a complex image denoising algorithm.

Time: 07.08.2017, 11:00
Location: Room 324 in the Fraunhofer IGD, Fraunhoferstrasse 5, S3|05
Speaker: Philipp Schneider (Advisor: Tatiana von Landesberger)
Title: "Visuelle Analyse von Zufallskaskaden in Netzwerken" (Bachelor thesis)

Time: 07.08.2017, 10:00
Location: Room 324 in the Fraunhofer IGD, Fraunhoferstrasse 5, S3|05
Speaker: Axel Platzwahl (Advisor: Felix Brodkorb)
Title: "Interaktives Erzeugen von dynamischen geographischen Netzwerken" (Bachelor thesis)
Abstract: Graphen finden in den unterschiedlichsten Gebieten Anwendung. Neben vielen anderen Ei- genschaften besitzen diese auch oft einen geographischen Kontext.
Beispielsweise Computer- netzwerke, bei denen die einzelnen Computer über geographische Standorte verfügen. Diese Besonderheit ist keinesfalls unwichtig, da sie zusätzliche Informationen bergen kann. Eben- falls ist es sehr wahrscheinlich, dass sich diese Netzwerke auch über die Zeit entwickeln. Zur genaueren Analyse der Netzwerke werden folglich einige Daten benötigt, welche aber nur in geringen Mengen frei zur Verfügung stehen. Genauso werden auch größere, vielfältigere Da- tenmengen zu Test- oder Simulationszwecken gebraucht, die kaum zu finden sind. Deswegen bedarf es an dieser Stelle an Wegen zur künstlichen Generierung dieser Daten.
In dieser Arbeit wird ein interaktiv, visuelles Generatormodel zur Erzeugung von dynamisch geographischen Netzwerken präsentiert. Dabei wird der Benutzer beim Steuern des Generie- rungsprozesses durch einen Wizard geführt, um die Bedienung zu erleichtern. Der eigentliche Algorithmus hinter dem Generator arbeitet zum Teil auf Power-Law Basis sowie auf weiteren, in dieser Arbeit präsentierten analytischen Ergebnissen.

Time: 18.07.2017, 14:00
Location: Room 324 in the Fraunhofer IGD, Fraunhoferstrasse 5, S3|05
Speaker: Patrick Rossnagel (Advisor: Tatiana von Landesberger)
Title: "Visual football data analysis using different perspectives in 3D and interactive what-if scenarios" (Master thesis)

GRIS Kolloquium

Time: 13.07.2016, 10:30
Location: Room 072 in the Fraunhofer IGD, Fraunhoferstrasse 5, S3|05
Speaker: Prof. Dr. Tobias Schreck
Institute for Computer Graphics and Knowledge Visualization
Graz University of Technology, Austria
Title: "Visual Analytics Support for Searching and Comparing Patterns in Scatter Plots"
Abstract:The Scatter Plot is a classic, well-known technique in data visualization, and it has been applied in many domains since a long time effectively. Based on mapping input data to marks in a 2D graph, it supports analysts to perceive correlations, clusters, outliers, data density and further patterns in data. Initially defined for two -dimensional data, extensions to the visual and interaction design of Scatter Plots allow to increase the dimensionality of the data to be explored. Hence, it has become an important baseline technique for many practical data analysis tasks.
We first introduce the classic Scatter Plot technique and some important extensions proposed in the literature. We then present several of our own previous works on Visual Analytics support for data analysis with this technique. These comprise approaches for feature-based interactive search for globally and locally similar patterns in large data sets, as well as interactive modeling of local data patterns. Further, we discuss approaches for automatically computing potentially interesting patterns in subspaces of higher-dimensional data spaces, and options to visualize these using Scatter Plots. We conclude the talk by discussing options for future work, including user-adaptive interest measures and novel interaction modalities.
Bio: Tobias Schreck is a Professor with the Institute for Computer Graphics and Knowledge Visualization at Graz University of Technology, Austria. Between 2011 and 2015, he was an Assistant Professor with the Data Analysis and Visualization Group at University of Konstanz, Germany. Between 2007 and 2011 he was a Postdoc researcher and head of a junior research group on Visual Search and Analysis with Technische Universität Darmstadt, Germany. Tobias Schreck obtained a PhD in Computer Science in 2006 from the University of Konstanz. He works in the areas of Visual Analytics, 3D Object Retrieval, and Digital Libraries. His research interests include visual search and analysis in time-oriented, high-dimensional and 3D object data, with applications in data analysis, multimedia retrieval, and cultural heritage. He serves as a paper co-chair for the IEEE VIS Conference on Visual Analytics Science and Technology (VAST) in 2017. He previously has served as co-chair for Posters, Workshops and Panels for IEEE VIS, as well as a co-organizer for the Eurographics Workshop on 3D Object Retrieval.

Thesis presentations

Time: 07.07.2017, 10:00
Location: Room 140 in the Fraunhofer IGD, Fraunhoferstrasse 5, S3|05
Speaker: Kristiyan Dimitrov (Advisor: Naser Damer)
Title: "Exploring deep multi-biometric fusion" (Maste thesis)
Abstract: I would like to invite you to the Master thesis presentation of Kristiyan Dimitrov on Friday 7. July at 10:00, room 140. His work is titled "Exploring deep multi-biometric fusion". He investigate the sanity of using deep learning to create multi-biometric representations by investigating different deep fusion possibilities and comparing them to the more common score-level fusion.

Time: 06.07.2017, 14:00
Location: Room 074 in the Fraunhofer IGD, Fraunhoferstrasse 5, S3|05
Speaker: Maximilian von Buelow (Advisor: Stefan Guthe)
Title: "Geometry and Attribute Compression of Triangle Meshes" (Bachelor thesis)
Abstract: Triangle meshes are used in various fields of applications and are able to consume a voluminous amount of space due to their sheer size and redundancies caused by common formats. Compressing connectivity and attributes of these triangle meshes decreases the storage consumption, thus making transmissions more efficient. I present in this thesis a compression approach using Arithmetic Coding that predicts attributes and only stores differences to the predictions, together with minimal connectivity information. It is applicable for arbitrary triangle meshes and compresses to use both of their connectivity and attributes with no loss of information outside of re-ordering the triangles. My approach achieves a compression rate of approximately 3.50:1, compared to the original representations and compresses in the majority of cases with rates between 1.20:1 to 1.80:1, compared to GZIP.


Habilitation TvL

Time: 03.07.2017, 12:45
Speaker: Frau Dr.-Ing. Tatiana Landesberger von Antburg
Titel: "Innovative Ansätze in der visuellen Datenanalyse"
more »

Thesis presentations

Time: 03.07.2017, 10:00
Location: Room 011 in the Fraunhofer IGD, Fraunhoferstrasse 5, S3|05
Speaker: Sven VetterAdvisor: Arjan Kuijper)
Title: "Depth Image Based Composition in Distributed Rendering Environments" (Master thesis)
Abstract: In dieser Arbeit wird ein auf Depth Image Based Rendering (DIBR) aufbauender Ansatz vorgestellt, der die flüssige Darstellung von Szenen auch auf leistungsschwächeren Geräten ermöglichen soll. Grundlage hierfür ist ein Client-Server Ansatz, bei dem der Server auf Anfrage Bilder zur Verfügung stellt, die der Client mit Hilfe von DIBR an seine Bedürfnisse anpasst. Außerdem wird ein Kameraprädiktor verwendet, um die Anfragen des Client zu optimieren. Die Qualität der erstellten Bilder wird mit Hilfe von drei verschiedenen Simulationen untersucht.

Guest lecture

Time: 30.06.2017, 10:00
Location: Room 074 in the Fraunhofer IGD, Fraunhoferstrasse 5, S3|05
Speaker: Prof. Dr. Peter Wonka
King Abdullah University of Science and Technology (KAUST)
Title: "Integer Programming for Layout Problems"
Abstract: In this talk, I will give an introduction to Integer Programming (IP) and show how we used IP in recent research projects. The projects range from problem formulations in visualization to urban modeling. The talk will focus mainly on the modeling side of integer programming, i.e. how to formulate a given problem as integer program.
Bio: Peter Wonka is Associate Director of the Visual Computing center (VCC) at King Abdullah University of Science and Technology (KAUST) and Professor in the Computer Science program. Peter Wonka received his doctorate from the Technical University of Vienna in computer science. Additionally, he received a Masters of Science in Urban Planning from the same institution. After his PhD, Dr. Wonka worked as postdoctoral researcher at the Georgia Institute of Technology and as faculty at Arizona State University. His research interests include various topics in computer graphics, visualization, remote sensing, computer vision, image processing, machine learning, and data mining.


Time: 29.06.2017, 11:00 - 12:00
Location: Room 072 in the Fraunhofer IGD, Fraunhoferstrasse 5, S3|05
Speaker: Anirban Mukhopadhyay
Junior Group Leader GRIS, TU Darmstadt
Title: "Deep Vision for Minimally Invasive Surgery"

Habilitation Lecture

Time: 22.06.2017, 16:45
Location: TU Darmstadt, Deoartment of Computer Science, Hochschulstraße 10, 64289 Darmstadt, Piloty-Building: S2|02, Raum C110
Speaker: Dr.-Ing. Tatiana Landesberger von Antburg
Title of Habilitation Lecture: ”Tree Visualization”
Title of Habilitation Thesis: ”Visual Data Comparison”

Thesis presentations

Time: 06.06.2017, 14:00
Location: Room 324 in the Fraunhofer IGD, Fraunhoferstrasse 5, S3|05
Speaker: Alexander Geurts (Advisor: Johannes Fauser)
Title: "Identifikation und Visualisierung unzugänglicher Zugangswege in nicht-holonomer Pfadplanung" (Master thesis)
Abstract: This work is about the identification and visualization of inaccessible regions for a medical operation at the human ear which uses a curved drill. In the context of non-holonomic path planning, two approaches are presented that try to identify narrow regions which are too small to be passed by the drill. The first approach uses spheres to explore the free space and tries to identify the size of the regions using the size of the spheres in it. This approach however has a couple of problems with the consistent identification of these regions. This is why another approach has been tested that provided good results. This one uses the voronoi diagram to search the graphical scene and identify the inaccessible regions. It is able to identify and visualize these regions on a consistent basis. Finally, a way to visualize the curvature of the drill is being presented. Therefor an approach by Xu et al. is being extended to visualize all realistically accessible points from a position and orientation.

Time: 24.05.2017, 11:30
Location: Room 072 in the Fraunhofer IGD, Fraunhoferstrasse 5, S3|05
Speaker: Sudesh Mirashi (Advisor: Stefan Wesarg)
Title: "Model-based segmentation of the teeth in panoramic radiograph images" (Master thesis)

Time: 18.05.2017, 09:00
Location: Room 324 in the Fraunhofer IGD, Fraunhoferstrasse 5, S3|05
Speaker: Felix Bernhard (Advisor: Johannes Fauser)
Title: "Nichtholonome Pfadplanung im Felsenbein unter Verwendung des Bevel-Tip-RRT" (Master thesis)
Abstract: Diese Masterarbeit behandelt zwei Schwerpunkte. Im ersten Teil geht es um die nichtholonome Pfadplanung im Felsenbein unter Verwendung des Bevel-Tip-RRT. Dieser neu entwickelte Bevel-Tip-RRT Algorithmus basiert auf Elementen der Veröffentlichungen von Xu et al [13] und Karaman et al [5] und kombiniert diese Ansätze, sodass sie für den Anwendungsfall der Bohrung mittels eines Roboters verwendet werden können. Diese Implementierung des Algorithmus wird mit dem existierenden Spline-Based-RRT Algorithmus nach Yang et al [14] experimentell verglichen, um die jeweiligen Vor- und Nachteile darstellen zu können.
Der zweite Teil dieser Arbeit befasst sich mit der Entwicklung eines mathematischen Modells für den verwendeten Bohrmolch-Roboter. Dieses Robotermodell wird durch geometrische Primitive approximiert und mit Hilfe von Denavit-Hartenberg Transformationen beschrieben. Des Weiteren wird eine Theorie zur Fortbewegung des Roboters im Felsenbein entwickelt, welche auf den Denavit-Hartenberg Transformationen des Modells beruht. Zur Überprüfung des Modells wird dieses auf sein Verhalten während der Fortbewegung getestet. Daraus wird der Schluss gezogen, ob dieses Modell mit den entwickelten Pfadplanungsalgorithmen verwendet werden kann.

Time: 15.05.2017, 15:15
Location: Room 324 in the Fraunhofer IGD, Fraunhoferstrasse 5, S3|05
Speaker: Benedikt Rohn (Advisor: Davis Kügler)
Title: "Kalibration von EMT Sensorpositionen" (Master thesis)
Abstract: Zur Durchführung von Operationen an dem Schädelknochen der Otobasis werden verschiedene minimal- invasive Ansätze zum Erreichen des Operationsortes untersucht. Für Nichtlineare Bohrpfade sind integrierte Mikro-Bohrsysteme notwendig. Zur Steuerung dieser Bohrsysteme muss die Lage (Position und Ausrichtung) kontinuierlich gemessen werden. Die Umrechung der mittels eines elektromagnetischen Trackers ermittelten Lage des Sensors im Roboter muss dazu mittels einer Kalibration bestimmt werden.
Auf Basis der Abtastung verschiedener Oberflächen des Bohrsystem-Prototypen wurden ein Kalibrations- protokoll, Hilfswerkzeuge zur Kalibration und ein Algorithmus zur Berechnung der geforderten Transformation entwickelt und ausgewertet. Für die Kalibration wurden mittlere Positionsfehler <0,1mm (0,4% der Haupt- dimension) und mittlere Winkelfehler <10° erreicht. Bei größerem Rauscheinfluss (SNR<10) nimmt der Fehler erheblich zu (1,7mm / 7% der Hauptdimension).

Dirk Bartz Prize 2017

(in German)

Auszeichnung: Dr. Jürgen Bernard erhält Dirk Bartz Preis for Visual Computing in Medicine

Dr. Jürgen Bernard wurde von der Eurographics Association (EG) mit dem Dirk Bartz Preis 2017 ausgezeichnet.
Der Preis wird an herausragende Leistungen im Bereich des Visual Computing vergeben, deren Ergebnisse sich durch ihre besondere Nützlichkeit in medizinischen Anwendungen auszeichnen.
Gemeinsam mit Prof. Kohlhammer und Dr. Thorsten May (beide Fraunhofer IGD) schaut J. Bernard auf eine erfolgreiche Kollaboration mit den Medizinern Prof. Thorsten Schlomm und Dr. Dirk Pehrke (beide Universitätsklinikum Hamburg-Eppendorf – UKE) zurück.
Ergebnis der Zusammenarbeit ist ein visuell-interaktives System zur Analyse von Patientenhistorien, am Beispiel vom Prostatakrebsleiden.
Mit dem System sind Ärzte nun in der Lage, Wissen aus großen Mengen von Patientenhistorien zu explorieren – und damit sowohl die Forschung als auch die Therapierung von Prostatakrebs zu verbessern.
Das System zeigt Signale zur Frühwarnung auf und eignet sich zudem zur Zusammenstellung und Auswertung von Kohorten für klinische Studien.
Diese komplexen Vorgänge in der Datenanalyse können nun direkt von den Ärzten in nur wenigen Minuten absolviert werden – ein Vorgang der zuvor Tage oder gar Wochen in Anspruch nahm.

Dr. Jürgen Bernard nahm den Preis am Freitag, den 28.04.2017 in Lyon auf der EuroGraphics Konferenz entgegen. Organisator und Stifter des Preises ist die Eurographics Association.


Thesis presentations

Time: 25.04.2017, 15:00
Location: Room 220 in the Fraunhofer IGD, Fraunhoferstrasse 5, S3|05
Speaker: Marija Schufrin (Advisor: Arjan Kuijper)
Title: "Designstudie und Entwicklung von Konzepten zur visuellen Trendanalyse für mobile Umgebungen" (Master thesis)
Abstract: Die immer schneller wachsenden Datenmengen in nahezu allen Bereichen des Lebens führen zu dem verstärkten Bestreben, die darin enthaltenen Informationen zu gewinnen und nützlich einzusetzen. Die menschlichen Kapazitäten im Hinblick auf die kognitive Verarbeitung von Informationen sind jedoch begrenzt. Daher werden Techniken und Methoden entwickelt, um den Menschen bei dieser Aufgabe zu unterstützen. Eine Ausprägung der Informationssuche in Datenmengen ist die analytische Untersuchung der Daten. Dazu gehört unter anderem das Erkennen von Trends und Mustern. Visual Analytics hat das Ziel, den Menschen durch eine Kombination aus Informationsvisualisierungen und automatisierten Analyseverfahren bei diesem Prozess zu unterstützen. Die meiste Forschung auf diesem Gebiet bezieht sich bislang überwiegend auf Computer und Laptops. Durch die rasante Verbreitung von mobilen Geräten in den letzten Jahren, entstand die Notwendigkeit, die Methoden und Techniken auch im Hinblick auf den Einsatz auf mobilen Geräten zu untersuchen. Die mobile Umgebung bringt neue Herausforderungen aber auch Möglichkeiten mit sich.
Am Fraunhofer IGD in Darmstadt wurde eine Software zur visuellen Trendanalyse auf Basis von digitalen Publikationsdatenbanken entwickelt. Ziel ist es, durch die Visualisierung der darin zur Verfügung stehenden Informationen, aufkommende und verschwindende Trends zu erkennen. Diese Software ist allerdings für die Benutzung auf einem Desktop ausgelegt. Im Rahmen dieser Arbeit wurde daher eine Designstudie durchgeführt, um die Designmöglichkeiten zur Umsetzung der Software auf mobilen Geräten zu untersuchen.
Dazu wurde ein Modell der Zielgruppe (Entscheidungsträger) erstellt, welches aus drei Eigenschaften besteht. Eine dieser Eigenschaften besteht aus zwei Merkmalen – Bauchgefühl und Verstand – welche als gegensätzliche Aspekte identifiziert wurden, deren Vorteile jedoch jeweils stark vom Kontext abhängen. Gerade bei mobilen Umgebungen hat der sich stets verändernde Kontext einen starken Einfluss auf die Benutzererfahrung. Adaptive Mechanismen mit Bezug auf die mentale Verfassung des Benutzers (z.B. geteilte Aufmerksamkeit) können große Vorteile im Hinblick auf die positive Nutzererfahrung mit sich bringen.
Im Rahmen dieser Designstudie wurden drei Designs entwickelt und untersucht, die jeweils das Bauchgefühl, den Verstand oder eine Kombination aus beiden in den Fokus stellen. Die Konzepte wurden prototypisch implementiert und im Rahmen eines kleinen kontrollierten Experiments evaluiert.

Time: 25.04.2017, 14:00
Location: Room 324 in the Fraunhofer IGD, Fraunhoferstrasse 5, S3|05
Speaker: Kenten Fina (Advisor: Roman Getto)
Title: "Automated detection of significant parameters in procedural 3D models" (Bachelor thesis)
Abstract: This bachelor thesis present an approach to automatically detect significant parameters in a procedural model. For the distinction of significant and insignificant parameters we present both a static version and a method using machine learning. In the process parameters are grouped, which represent symmetries or other relations in the model. Additionally we allow the user to adapt the selection of significant parameters to his needs. For this purpose we support the user by visualizing the changes of a parameter. Furthermore a hierarchical arrangement of the parameters is done to give the user an overview of all design possibilities. Subsequently, we show how ranges for the selected parameters can be calculated, which retain the object.

Time: 18.04.2017, 16:30
Location: Room 324 in the Fraunhofer IGD, Fraunhoferstrasse 5, S3|05
Speaker: Lukas Baumann (Advisor: Tatiana von Landesberger)
Title: "Visual Tourist Guide extracted from Flicker Images" (Bachelor thesis)

Time: 18.04.2017, 15:30
Location: Room 324 in the Fraunhofer IGD, Fraunhoferstrasse 5, S3|05
Speaker: Axel Patzwahl (Advisor: Felix Brodkorb)
Title: "Generieren von zeitabhängigen Geo-Netzwerken" (Intermediate presentation Bachelor thesis)

Time: 18.04.2017, 15:00
Location: Room 242 in the Fraunhofer IGD, Fraunhoferstrasse 5, S3|05
Speaker: Katrin Hartwig (Advisor: Christina Oyarzun Laura)
Title: "Automatische Segmentierung der Nasenscheidewand auf Basis von computertomographischen Bilddaten" (Bachelor thesis)
Abstract: Im Kontext der Nasenheilkunde spielt die Segmentierung der Nasenscheidewand anhand von computertomographischen Bildern eine wichtige Rolle für Diagnose, präoperative Planung und Behandlung. Aktuelle Verfahren stützen sich zumeist auf manuelle Segmentierung, welche für den medizinischen Alltag zu zeitintensiv ist. In der vorliegenden Bachelorarbeit wurde daher ein vollautomatisches Verfahren entwickelt, welches an die anatomischen Besonderheiten und computationellen Herausforderungen der Nasenscheidewand adaptiert ist. Dabei wurde eine Kombination aus Musterdetektion und Slice-based Propagation angewandt.
Die Evaluation anhand von 19 Datensätzen zeigt mit durchschnittlich 0,78 Sekunden pro CT-Bild eine deutliche Beschleunigung im Vergleich zu manuellen Verfahren und erzielt auch im Bezug auf die Genauigkeit der Segmentierung mit einem durchschnittlichen DSC-Wert von 0,8665 annehmbare Ergebnisse.

Time: 18.04.2017, 14:00
Location: Room 048 in the Fraunhofer IGD, Fraunhoferstrasse 5, S3|05
Speaker: Christian Ritter (Advisor: Jürgen Bernard)
Title: "Personalized Music Classification and Feature Creation based on Visual-Interactive Learning" (Bachelor thesis)
Abstract: : Mit den Ergebnissen seiner Bachelorarbeit eröffnet Christian Ritter den Nutzern privater Musikkollektionen neue Möglichkeiten. Er geht davon aus das Musik von verschiedenen Menschen verschieden wahrgenommen, bewertet, sortiert und gruppiert (klassifiziert) wird. Zudem ist die Klassifikation von Musikstücken stimmungsabgängig und über die Zeit (Erfahrung) durchaus veränderlich. Allerdings stellt Christian fest, das gegenwärtige Musiksysteme den individuellen Nutzerwunsch bei der Klassifikation von Musik in der Regel kaum bzw. gar nicht berücksichtigen. Vielmehr existiert häufig eine fest vorgegebene, unveränderliche Standardklassifikation, die vom System für die Allgemeinheit von Musikhörern herangezogen wird um zum Beispiel die Erstellung von Playlisten zu ermöglichen.
Angelehnt an das Konzept "Personal Analytics" hat es sich Christian zur Aufgabe gemacht die Funktionalität gegenwärtiger Systeme in Sachen Personifizierung und Integration von Nutzerinteresse zu übertreffen.
Um diese Schwierigkeit zu meistern, bedient sich Christian an einer Reihe von Techniken aus dem Bereich des Machine Learning (ML) und Music Information Retrieval (MIR). Darauf aufbauend präsentiert er zwei neuartige visuell-interaktive Systeme welche Nutzer in die Lage versetzen Musik visuell-interaktiv und nach eigenem Nutzerwunsch zu klassifizieren. Visual Analytics Techniken ermöglichen zudem die visuelle Validierung und Optimierung der Musikklassifikation. Das zweite visuell-Interaktive System bietet Nutzern die Möglichkeit Einfluss auf den zugrundeliegenden Musik-Featurevektor für die Klassifikation zu nehmen. Mit Hilfe eines visuellen Interface können Nutzer interessante Musikpatterns (wie etwa einen Beat, Clap, Akkord, Takt, etc.) in individuell ausgewählten Musikstücken definieren, auf deren Basis das System dann automatisch einen personifizieren Musik-Featurevektor generiert, der für die Klassifkation verwendet werden kann.
In einem seiner Nutzungsszenarien zeigt Christian zum Beispiel auf, wie ein Ensemble-Classifier erfolgreich lernt Livemusik von Studiomusik zu unterscheiden.

Time: 18.04.2017, 10:50
Location: Room 324 in the Fraunhofer IGD, Fraunhoferstrasse 5, S3|05
Speaker: Evgheni Croitor(Advisor: Johannes Fauser)
Title: "Automatische Pfadplanung für nicht-lineare Bohrkanäle im Felsenbein" (Master thesis)
Abstract: Die vorliegende Arbeit beschäftigt sich mit der Aufgabe, einen diskreten Algorithmus zu entwickeln, der automatisch einen optimalen, nicht-linearen Bohrkanal zum Innenohr findet. Hierfür präsentiere ich zwei unterschiedliche Ansätze: einen auf Basis des Voronoi-Diagramms und einen anderen mit Hilfe der Motion primitives. Es lässt sich zeigen, dass der auf Motion primitives basierende Ansatz über gewisse Vorteile verfügt, die aus der gegebenen Problemstellung resultieren. Schließlich wird dieser mit dem aktuell verwendeten RRT-Algorithmus verglichen und die empirischen Ergebnisse ausgewertet.

Time: 18.04.2017, 10:00
Location: Room 324 in the Fraunhofer IGD, Fraunhoferstrasse 5, S3|05
Speaker: Wei Hung Hsu (Advisor: Johannes Fauser)
Title: "Felsenbeinsegmentierung" (Bachelor thesis)
Abstract: Für eine Operationsplanung im Mittel- und Innenohrbereich ist die Konstruktion von 3D Modellen der inneren Strukturen aus Schnittbildern, die durch bildgebende Verfahren erzeugt werden, oftmals ein wichtiger Bestandteil. Für die Konstruktion ist es allerdings notwendig, die Strukturen aus den Bildern korrekt zu segmentieren und zu klassifizieren. In dieser Arbeit wird ein bildbasiertes Segmentierungsverfahren für das Felsenbein präsentiert. Die Eingabedaten sind dreidimensionale Grauwertbilder der seitlichen Schädelbasis, die durch die Computertomografie erfasst wurden. Die Segmentierung beinhaltet neben dem knöchernen Felsenbein auch die natürlich vorhandenen inneren Lufteinschlüsse des Mastoids. Das Verfahren besteht aus einer Kombination von bildbasierten Segmentierungsalgorithmen. Weiterhin lässt sich das Verfahren in 4 Phasen unterteilen: (1) die Vorverarbeitung, in der das Eingabebild für die weitere Verarbeitung verbessert wird, (2) die Segmentierung des knöchernen Bestandteils und der Lufteinschlüsse im Felsenbein, (3) die Segmentierung der Paukenhöhle zusammen mit der Eustachischen Röhre und schließlich (4) die Nachverarbeitung, in der die einzelnen Segmentierungen miteinander verknüpft werden. Die Ergebnisse weisen darauf hin, dass die ledigliche Verwendung von bildbasierten Algorithmen ohne manuelle Interaktion zu einer groben Segmentierung der inneren Lufteinschlüsse führen kann.

Time: 02.03.2017, 13:00
Location: Room 242 in the Fraunhofer IGD, Fraunhoferstrasse 5, S3|05
Speaker: Martin Majewski (Advisor: Andreas Braun)
Title: "3D-printed electrodes for electric field sensing technologies" (Master thesis)
Abstract: Elektrische Feldsensorik, sowie kapazitive Sensorik sind seit mehr als einem Jahrhundert ein intensiv betriebenes Forschungsfeld. Erschwingliche Rapid-Prototyping Technologien, welche immer mehr an Beliebtheit gewinnen, wie z.B Kleinstrechner für die Erstellung elektronischer Schaltungen und vor allem der 3D-Druck, welcher auf "fused filament fabrication" basiert, eröffnen neue Möglichkeiten. Der 3D-Druck fördert die Ambitionen im Bereich individuell erstellter Objekte mit vollständig und unauffällig integrierter Elektronik. Elektrisch leitende Materialien für den 3D-Druck können in dem Zusammenhang dazu genutzt werden Elektroden für elektrische Feldsensorik zu erstellen. Diese Elektroden können als Bestandteil des Gesamtobjekts direkt in jenes hineingedruckt werden. Allerdings wurde im Zuge verwandter Arbeiten bisher keine Untersuchung der 3D-gedruckten Elektroden durchgeführt, welche ihre Messleistung und die Kosten in Bezug zum verwendeten leitenden 3D-Druckmaterial (Filament), der 3D-Druck-Konfiguration und der ausgewählten 3D-Druck-Muster stellt. Diese Thesis stellt einen ersten Ansatz für den Einblick dar, welcher die Wechselwirkung zwischen den gewählten 3D-Druck-Parametern und der Gesamtmessleistung dieser Elektroden aufgreift. Dafür wurden 30 Elektroden 3D-gedruckt und deren Messleistung bezüglich einer Kupfer- und einer Placebo-Elektrode evaluiert. Die Evaluation wurde mithilfe eines eigens konstruierten Messaufbaus, dem CapLiper, durchgeführt, welcher wiederum auf Funktionalität und Verlässlichkeit der Messungen evaluiert wurde. Die Resultate der Elektrodenevaluation zeigen, dass 3D-gedruckte Elektroden mit der Messleistung der Kupferelektrode mithalten können, einige diese sogar übersteigen. Unter Bezug dieser Resultate, als auch aus den Erfahrungen, welche durch den Bau zweier unterschiedlicher Prototypen gewonnen wurden, führt diese Thesis Empfehlungen zur Vorgehensweise beim 3D-Druck von Elektroden auf. Weiterhin wird ein Ausblick auf potentielle zukünftige Arbeiten gegeben.

Time: 22.02.2017, 16:00
Location: Room 324 in the Fraunhofer IGD, Fraunhoferstrasse 5, S3|05
Speaker: Volker Knauthe (Advisor: Marcel Wunderlich)
Title: "Visual Interface For Train Trip Planning Considering Expected Train Delays" (Bachelor thesis)
Abstract: Planning long distance train journeys may require travellers to choose between a lot of different connections, which may differ in various details, such as the total travel duration, waiting times in railway stations, or the possibility to transport bicycles. Expected delays, expected transfer times and possible alternatives for missed trains may additionally make a connection more or less favourable regarding the given options. The Thesis is based on the solution proposed in Visual Analysis of Train Schedules Regarding Expected Delays and User Preferences [Wun16] by Marcel Wunderlich, which focussed on the visualization of possible delays, alternatives and the demand for user preferences. In the following paper we introduce a web application, that contains the given results and adds new features. A new visualization for transfer times, should enable Users to decide if a given connection is suitable for them. This includes possible delays and the resulting reduced possible transfer times. Furthermore different possibilities to sort and/or filter connections are introduced. Time sliders enable the User to search connections regarding precise travel time restrictions, i.e. appointments. Preferences enable the User to decide if he wants to travel with the least amount of risk or to choose potential delayed connections. Additionally the visualized connections may now include multiple alternatives at every critical point, where an alternative is possible. This may lead to new decisions, as some connections have a broader all together flexibility to reach your destination. Detailed informations on demand allow the comparison between similar connections and may provide every information, that is given and possibly needed, to choose the best suited connection for Users. If these features pose an advantage and if the System is usable was tested and evaluated with a user study. Two groups were tested, where one had the options to sort, filter and access to detailed informations, while the other had no addi- tional features but the visualization. The results show, that not everyone who could actually used features. But Users who used the features had overall faster decision times and made less mistakes finding optimal connections. Furthermore we could observe, that multiple alternatives and the transfer time visualization proved to be easy and fast to understand.

Time: 14.02.2017, 10:15
Location: Room 074 in the Fraunhofer IGD, Fraunhoferstrasse 5, S3|05
Speaker: Daniel Kauth (Advisor: Prof. Dr.-Ing. Michael Goesele)
Title: "A Novel Approach to Surface Refinement using an Implicit Function" (Master thesis)
Abstract: The successful algorithms for reconstructing surfaces from unorganized point clouds in difficult outdoor environments mainly fall into two categories. The first one is based on a volumetric segmentation of space into parts which are inside and outside of the object to be reconstructed. Such a segmentation can be found in a way that globally minimizes the violation of visibility constraints. Thanks to the global optimization, these methods are very robust against outliers and are able to create approximate geometry even for sparsely sampled areas. However, they interpolate the input points, which makes them unable to produce highly detailed geometry by averaging out sample noise. The algorithms of the second category use the sample positions and additional attributes like normals, scale values and confidences. They define a global function which implicitly represents the surface. The surface is extracted using a contouring algorithm. These methods are able to produce highly detailed surfaces by averaging sample contributions in a local neighborhood. However, they are generally less robust against outliers and often are unable to reconstruct surfaces with small or no sample support. The goal of this thesis is to explore the possibility of combining aspects of algorithms of both categories to overcome their shortcomings. Specifically, we use a volumetric segmentation algorithm to create an initial surface which is mostly unaffected by outliers and includes sparsely sampled areas. We propose a novel refinement algorithm, based on an implicit function, to improve the reconstruction quality of the initial surface. Our method is based on moving the vertices of the initial mesh onto the implicitly defined surface. We also achieve an adaptive mesh resolution by subdividing all faces which do not approximate the implicit surface well enough. We evaluate our algorithm on multiple challenging datasets and show that it is able to create highly detailed and complete reconstructions while being robust against outliers.

Time: 14.02.2017, 09:30
Location: Room 074 in the Fraunhofer IGD, Fraunhoferstrasse 5, S3|05
Speaker: Daniel Thul (Advisors: Prof. Dr.-Ing. Michael Goesele, Prof. D.Sc. (Tech.) Jaakko Lehtinen)
Title: "Deep Appearance – Synthesizing Materials with Neural Networks" (Master thesis)
Abstract: Materials are an important aspect in modern fields such as movies and games. Capturing the essence of material appearance allows us to make virtual material creation and modification much more straightforward. We present a method for material texture synthesis using generative adversarial networks (GANs). Unlike previous methods we do not employ activation maximization to drive the generator to an optimal solution but use a loss function that compares activations between a given exemplar and the output of the generator. This allows us to accurately specify the appearance of the image we generate as opposed to activation maximization methods that only allow you to choose the class you want to generate. We show that it is possible to generate tileable images with what we call circular transposed convolutions. Additionally we present normalized transposed convolution that mitigates the checkerboard artifacts one can observe when using strided transposed convolutions.

Time: 27.01.2017, 10:00
Location: Room 073 in the Fraunhofer IGD, Fraunhoferstrasse 5, S3|05
Speaker: Chinara Mammadova (Advisor: Helmi Ben Hmida)
Title: "New approach for optimizing usage of situation recognition algorithms within IoT domains" (Master thesis)
Abstract: Over the past few years technological advancements have supported the growth of the Internet of Things (IoT). The Internet of Things consists of (smart) objects embedded with sensors, actuators and controllers. These objects are connected to the Internet and are able to communicate with each other. The interconnection and communication of objects enable the creation of different application domains within the Internet of Things. Smart living is one of the major application areas for the Internet of Things.
Sensors, actuators and controllers in a smart living environment (e.g. smart homes) are deployed anywhere; on objects or even on persons. As sensors have the capability to sense the environment, they can be used to collect useful information on location, motion, temperature, humidity, light, etc. Actuators can perform different actions based on data gathered from sensors, and controllers can process that data. Real-time situation awareness is one of the key tasks in a smart living environment. Real-time recognition of situations is especially important in ambient assisted living environments, where elderly or disabled people need support in their everyday lives. Recognition of situations in real-time enables immediate identification of critical situations and provides just-in-time assistance. To detect situations, data needs to be monitored, collected, analyzed and processed. Due to the increasing number of IoT connected devices, the amount of generated data is increasing too. Processing huge amounts of data is complex due to the inefficiency of continuously-running pattern/situation recognition algorithms, high requirement for processing capability and high frequency of the recognition process. Situation recognition algorithms must be executed constantly to handle the continuously generated data. For real-time recognition of situations in particular, these algorithms need to be executed permanently for all received data. The continuously-running recognition algorithms require high processing capabilities. The resource consumption of these algorithms is especially high when they are running on large sets of data. To overcome these problems there is a need for more intelligent approaches that are able to decide – based on target situation recognition purposes – which data is important and should be processed and which algorithm should be used to process this data.
This study proposes an approach for optimizing the usage of situation recognition algorithms in Internet of Things domains. The key idea of our approach is to select important data, based on situation recognition purposes, and to execute the situation recognition algorithms after all relevant data have been collected. The main advantage of our approach is that situation recognition algorithms will not be executed each time new data is received. This allows reduction of the frequency of execution of the situation recognition algorithms, thus saving computational resources, such as CPU, memory, storage, bandwidth and power. Another advantage of our approach is that it can be applied to recognize situations in real-time, which is useful for ambient assisted living environments. We apply the proposed approach to implement a use case scenario on top of the universAAL IoT platform, which is an open-source platform for the development of IoT solutions.

Time: 25.01.2017, 15:30
Location: Room 324 in the Fraunhofer IGD, Fraunhoferstrasse 5, S3|05
Speaker: Jannik Sehring (Advisor: David Kügler)
Title: "3D/2D Image Tracking in 2D Röntgenbildern mit Active Shape Model" (Bachelor thesis)
Abstract: Im Rahmen des MUKNO-Projekts wird ein Bohrroboter für minimalinvasive Eingriffe entwickelt. Für die Positionsbestimmung während der Operation stehen mehrere Verfahren zur Verfügung. Diese Ausarbeitung beschäftigt sich hierbei mit der Positionsbestimmung anhand von zweidimensionalen Röntgenbildern.
Dafür wird eine Segmentierung mittels Active Shape Model durchgeführt. Die optimierten Parameter des statistischen Modells werden hierbei für die Positionsbestimmung genutzt. Zusätzlich wird eine Volumenprojektion mittels des Verfahrens von Siddon-Jacobs umgesetzt. Dabei werden Digitally Reconstructed Radiographs erstellt, die den Bohrroboter und einen anatomischen Hintergrund enthalten. Diese Bilder werden für das Training und das Testen des Segmentierungsverfahrens verwendet.

Time: 17.01.2017, 10:00
Location: Room 103 in the Fraunhofer IGD, Fraunhoferstrasse 5, S3|05
Speaker: Benedikt Hiemenz (Advisor: Michel Krämer)
Title: "Authentication and Searchable Symmetric Encryption for Cloud-based Storage of Geospatial Data" (Master thesis)
Abstract: Das Auslagern von Daten und Prozessen in die Cloud ist ein wachsender Trend bei Unternehmen, welche sich hierdurch vor allem wirtschaftliche Vorteile versprechen. Weil die Cloud Infrastruktur oftmals von externen Dienstleister bereitgestellt wird, müssen Unternehmen ihre vertraulichen Daten in der Cloud schützen. Ein Beispiel dafür sind raumbezogene Daten, welche sensible Informationen über Stadtgebiete enthalten. Zur Absicherung solcher Daten präsentieren wir zwei Sicherheitserweiterungen für cloudbasierte Datenspeicher. Zuerst stellen wir ein Authentifizierungsverfahren vor, das Anwendungen eine sichere Identifizierung seiner Nutzer ermöglicht. Dabei erweitern wir tokenbasierte Verfahren um wesentliche Eigenschaften wie den Widerruf von Token und erlauben so die gezielte Sperrung von Benutzern. Gleichzeitig benötigt unser Verfahren nur minimale Informationen innerhalb der Anwendung und ist damit für den Einsatz in verteilten Systemen geeignet. Unsere Haupterweiterung behandelt die Durchsuchbarkeit von verschlüsselten, raumbezogenen Daten. Wir entwickeln ein Searchable Symmetric Encryption Verfahren bei dem ein Index für die Suche genutzt wird. Unser dynamisches System erlaubt es jederzeit dem Index neue Daten hinzuzufügen und zu löschen. Wir präsentieren mehrere Versionen unseres Verfahrens mit denen Nutzer verschiedene Sicherheitsstufen erreichen und davon abhängig auch die Performanz bestimmen. Alle Versionen unterstützen spezifische Suchkriterien wie die Anfrage anhand einer Bounding Box und erlauben darüber hinaus boolesche Ausdrücke. Zudem benötigt unser Verfahren nur kryptographische Schlüssel auf Client Seite und ist somit für den Einsatz auf mehreren Geräten ausgelegt. Die Evaluierung beider Sicherheitserweiterungen zeigt, dass diese geeignet sind für cloudbasierte Anwendungen aber die Suche in verschlüsselten Daten die Laufzeit beeinträchtigt.

Time: 19.12.2016, 10:00
Location: Room 072 in the Fraunhofer IGD, Fraunhoferstrasse 5, S3|05
Speaker: Martin Müller (Advisor: Tatiana von Landesberger)
Title: "Guidance zur Exploration von Entitätsgraphen aus Textkollektionen für Datenjournalismus" (Bachelor thesis)


Time: 22.11.2016, 15:20 - 16:50
Location Raum 074 im Fraunhofer IGD, Fraunhoferstrasse 5, S3|05
Speaker: Thierry Driver
Academic Programs Coordinator, Tableau Software
Title: "The Beautiful Science of Data Visualisation"
- Introduction to Data visualisation (including the concepts and best practices)
- Tableau's mission (what is Tableau, our products, some of our customers, etc.)
- Demo/hands-on training on Tableau

Thesis presentations

Time: 15.11.2016, 13:00
Location: Room 324 in the Fraunhofer IGD, Fraunhoferstrasse 5, S3|05
Speaker: Erik Spiegelberg (Advisor: Roman Getto)
Title: "Maschinelles Lernen von parametrischen Modellen zur Klassifizierung von 3D Objekten" (Master thesis)
Abstract: Das Ziel dieser Masterarbeit ist es, parametrische Modelle als Datenbasis für maschinelles Lernen zu untersuchen. Dabei wird untersucht wie gut diese Datenbasis zur Klassifikation von 3D Objekten geeignet ist. Als Ansatz für parametrische Modelle wird die Generative-Modelling-Language(GML) verwendet. Diese ermöglich die Erstellung von parametrischen 3D Modellen. Zunächst wird damit eine Auswahl an Objekten Modelliert und eine Vielzahl von Exemplaren produziert. Zur Transformation der Modelle in lernbare Daten werden die 3D Objekt Deskriptoren PANORAMA und DESIRE verglichen. Als Methoden des maschinellen Lernens werden Varianten von k-Means, SVM und ein Nearest-Neighbor Ansatz Verwendung. Es werden verschiedene Szenarien erstellt, welche die Eignung von GML und PANORAMA für machinelle Lernansätze zeigen. Bei der Auswertung werden verschiedene Teile des PANORAMA Deskriptors auf ihre Klassifizierungseffizienz untersucht. Die Ergebnisse Zeigen das maschinelles Lernen anhand von parametrisch konzeptionierten Modellen zur Klassifikation von realen Objekten geeignet ist.

Time: 14.11.2016, 12:00
Location: Room 140 in the Fraunhofer IGD, Fraunhoferstrasse 5, S3|05
Speaker: Anton Zeltser (Advisor: Arjan Kuijper)
Title: "Entwicklung und Evaluierung eines Systems zur bildbasierten Detektion von Fehlern in Stoffen" (Masterarbeit)
Abstract: Die Algorithmen aus den Bereichen Bildverarbeitung, Computer Vision und Maschinelles Lernen finden in der heutigen Zeit immer häufiger ihre Anwendung bei den Industrieprozessen. In vielen Bereichen der Technik und Industrie sind diese Algorithmen der wichtige Bestandteil des
Planungs- und Produktionsprozesses geworden. Insbesondere in dem Qualitätskontrollvorgang haben bildbasierte Verfahren eine größere Bedeutung. Diese Verfahren ermöglichen es, die Qualität der Produktion, entsprechend den Qualitätsanforderungen, automatisch und präzise zu überprüfen und mögliche Fehlerteile zu identifizieren.
Im Mittelpunkt dieser Masterarbeit steht die Schritt-für-Schritt Entwicklung und Analyse eines Systems (in Form eines Algorithmus) zur bildbasierten Detektion von Fehlern in Materialien. Zur Kontrolle werden gebrauchte Stoffstücke mit folgenden Defekten verwendet: Löcher, Risse und Silikonflecke. Eine Besonderheit bei der Erkennung liegt darin,dass die Materialien im Bild sehen so aus, als ob sie unabsichtlich auf einen Tisch geworfen worden sind. Die zu prüfenden Materialien kann nicht als eine aufgespannte (2-D) Ebene beschrieben werden. Auf diese Art können vorhandene Textilien im Bild neben Defekten verschiedene Merkmale, wie zum Beispiel Falten, innere und äußere Ränder, besitzen. Diese zusätzliche Merkmale könnten irrelevante Informationen bei der Suche nach Defekten beitragen. Um die irrelevante Information zu reduzieren, wird vorgeschlagen, dass ein vorhandenes Bild in Ausschnitte aufgeteilt wird. Damit lässt sich jeder Ausschnitt meistens mit einem Merkmal beschreiben. Diese Bildausschnitte werden als Inputdata für das System genutzt.
Ziel der Arbeit ist es dabei, zum einen, mit Hilfe von Algorithmen des Maschinellen Lernens ein System für die Detektion des Defekts in Stoffen aufzubauen, zum anderen, das Systems sollte auseinanderhalten, ob ein Stoffstück Verschleiß oder Silikon besitzt, oder, ob ein Stoffstück fehlerfrei mit Falten ist. Anhand der Ausschnitte wird eine Entscheidung über den Defekt im Bild getroffen.
Diese Arbeit zeigt, wie die Algorithmen Local binary patterns in Verbindung mit Klassifikationsverfahren Support Vector Machine für die Detektion der Defekte in Stoffen verwendet werden.

Time: 26.10.2016, 11:30
Location: Room 073 in the Fraunhofer IGD, Fraunhoferstrasse 5, S3|05
Speaker: Andreas Bartschat (Advisor: Arjan Kuijper)
Title: "Automatic Classification of Cornea Tissues for Autofocus Algorithms" (Master thesis)
Abstract: Corneal confocal microscopy (CCM) is a spreading technique for investigations of cellular structures in the human cornea. It is non-invasive and allows in vivo imaging of the different tissue layers in the cornea with high resolution. High expectations are currently placed on CCM, to allow rapid and detailed analysis of pathological alterations affecting the peripheral nerves that innervate the cornea, resulting not only in fast diagnosis, but also providing insights into the progress and severity of diseases like diabetes. For the fast and reliable imaging of the sub-basal nerve plexus (SNP), the layer with the highest density of nerves, the focus must be adapted to compensate anatomical layer irregularities and reversible folds.
This thesis analyses classification methods of the anatomical tissues surrounding the SNP, to find reliable and fast classification models, suitable for online focus adaptations of the microscope to the layer of interest based on the predicted tissue type of the current image. The proposed methods are able to achieve accuracies of more than 92% with a runtime of less than 10 ms per image, evaluated on datasets of more than 8000 images as well as successfully applied in studies for the reconstruction of large field of view images of the SNP.

Time: 14.10.2016, 13:00
Location: Room 048 in the Fraunhofer IGD, Fraunhoferstrasse 5, S3|05
Speaker: Manuel Wolf (Advisor: Samir Aroudj)
Title: "Floating Scale Multi-View Visibility Constraints for Surface Reconstruction" (Bachelor thesis)
Abstract: Wir beschreiben in dieser Arbeit eine neue Methode zur Filterung von Multi-Scale- Punktwolken anhand einer expliziten FreeSpace-Definition zur optischen Verbesserung der Rekonstruktionsergebnisse. Dabei steht die Punktwolke eingeordnet in einem Octree zur Verfügung, wobei jedes Sample die Kameras aus einem MVS-Schritt kennt, die es erzeugt haben. Anhand der Blätter des Baumes teilt unser Algorithmus den Szenenraum in FreeSpace ein. Dies geschieht anhand zweier Gewichtsfunktionen, die für jedes Blatt ausgewertet werden und daraus ein Occupancy-Wert berechnet wird. Überschreitet dieser einen Grenzwert, so wird das jeweilige Blatt als FreeSpace markiert. Alle Samples die im daraus resultierenden Freiraum liegen, werden anschließend zusammengetragen und gelöscht.

Time: 27.09.2016, 10:00
Location: Room 324 in the Fraunhofer IGD, Fraunhoferstrasse 5, S3|05
Speaker: Sarah Berkei (Advisor: Max Limper)
Title: "Effiziente und Vollautomatische Grobausrichtung für den Soll-Ist-Abgleich zwischen CAD-Modellen und Scandaten" (Master thesis)
Abstract: Diese Arbeit befasst sich mit der Entwicklung und anschließenden Evaluation eines globalen Registrierungsverfahren zum Soll-Ist-Abgleich zwischen Scandaten und einem CAD-Modell. Dabei liegt der Fokus auf einer effizienten Ausrichtung der Daten zueinander, die schneller als ein manuelles Vorgehen ist. Als Grundlage dient der „4-Point Congruent Sets“ Algorithmus, der die Transformation zwischen zwei Punktwolken zueinander berechnet. Im Rahmen dieser Arbeit wird er mit einem klassischen, auf den Anwendungsfall  der  Arbeit  angepassten, Ansatz  „RANSAC-based  DARCES“  evaluiert. Darüber  hinaus werden das „Largest Common Pointset“ und die „Hausdorff-Distanz“ als Bewertungskriterien verglichen,
um eine möglichst effiziente und genaue Registrierung zu erreichen. Um das Ergebnis der globalen Registrierung zu bewerten, wird der „Iterative Closest Point“ Algorithmus in die globale Registrierung integriert. Zu erkennen ist, dass 4PCS, vor allem auf strukturreichen Oberflächen, mit diesem Ansatz in durchschnittlich 5,5 Sekunden bis zu 74% zuverlässige Ergebnisse liefert. Auf strukturarmen Oberflächen kann mittels des „RANSAC-based DARCES“ bis 82% valide Ergebnisse erzielt werden, wobei die Zeit für die Berechnung mit den verwendeten Testmodellen durchschnittlich unter 4 Sekunden lag.

Time: 25.08.2016, 15:00
Location: Room 242 in the Fraunhofer IGD, Fraunhoferstrasse 5, S3|05
Speaker: Tim Alexander Bergmann (Advisor: Pavel Rojtberg)
Title: "Interaktive Echtzeit-Kalibierung" (Master thesis)
Abstract: Diese Arbeit beschäftigt sich mit der Entwicklung einer Heuristik, die unerfahrene Nutzer durch eine flexible Kamerakalibrierung leitet. Hierfür wird ein Qualitätsmaß basierend auf der Arbeit von Hartley und Zisserman hergeleitet. Dieses Qualitätsmaß wird verwendet, um Nutzern mit Hilfe von Vorschlägen für Kameraposen die Durchführung einer Kalibrierung zu vereinfachen. Durch diese Hilfestellungen gelingt es unerfahrenen Nutzern eine Kamera mit weniger Bildern, aber gleicher Qualität, bezogen auf den Reprojection Error, zu kalibrieren.

Time: 21.07.2016, 09:00
Location: Room 220 in the Fraunhofer IGD, Fraunhoferstrasse 5, S3|05
Speaker: Thu Huong Luu (Advisor: Johannes Mueller-Roemer)
Title: "Adaptives und hybrides SLAM für handgeführte RGBD-Kamers" (Master thesis)
Abstract: Mit der steigenden Beliebtheit von RGBD-Sensoren wurde viel Forschung im Bereich der Aufnahme und Rekonstruktion von dreidimensionalen Umgebungen mit Hilfe von solchen Sensoren betrieben. Für die Konstruktion muss das sogenannte Simultaneous Localization and Mapping (SLAM)-Problem gelöst werden. Die meisten RGBD-SLAM-Systeme verwenden hierbei den punktbasierten Iterative Closest Point (ICP)-Algorithmus. Auch wenn ICP ein gut untersuchter Algorithmus ist, so stößt er bei verrauschten Daten und besonders bei texturarmen Bereichen mit wenigen geometrischen Merkmalen, wie z.B. großen leeren Flächen, auf Probleme. Eine Option, diese Limitierung anzugehen, ist das zusätzliche Ausnutzen von Ebenen in der Szene, besonders da sie die häufigste Form in von Menschen erbauten Innenräumen und Außenanlagen sind.
Taguchi et al. [TJRF13] veröffentlichte 2013 die erste globale Registrierungsmethode, in welcher Punkt-zu-Punkt- und Ebene-zu-Ebene-Korrespondenzen zu einem echtzeitfähigen SLAM-System vereint werden. Kurz darauf folgte die Publikation von Ataer et al. [ACTRG13], welche zusätzlich ein Bewegungsvorhersage-Modell ausnutzt, um Korrespondenzen zu bestimmen. Ein Nachteil dieser Verfahren ist die hohe Verarbeitungszeit eines Registrierungsschrittes. Dieser bewirkt, dass die Verfahren nicht in der Lage sind, interaktive Rekonstruktionen durchzuführen.
Das Ziel dieser Arbeit ist die Implementierung eines SLAM-Algorithmus für handgeführte RGBDKameras, der sowohl Punkte, als auch Flächen zur Registrierung nutzt. Im Gegensatz zu bestehenden Verfahren wird in dieser Arbeit ein lokaler Registrierungsalgorithmus umgesetzt. Flächenmerkmale werden bevorzugt verwendet, da ihre Anzahl in Szenen signifikant geringer ist als die von Punkten. Das ermöglicht eine schnellere Korrespondenzsuche und Registrierung. Dem zugrundeliegenden RANSACbasierten Algorithmus reicht bereits eine minimale Anzahl an Korrespondenzen aus, um die Sensorpose zu bestimmen. Somit ist der Algorithmus in der Lage, die Registrierung auch in texturarmen Bereichen mit wenigen geometrischen Merkmalen durchzuführen, in denen Techniken, welche nur Punkte benutzen, scheitern. Des Weiteren ermöglicht der lokale Registrierungsansatz eine interaktive Nutzung, um dem Nutzer in Echtzeit Rückmeldung über den Registrierungsprozess zu geben. Zusätzlich implementierte Erweiterungen, welche die detektierten Flächeninformationen zur Geometriekorrektur ausnutzen, unterstützen den Registrierungsvorgang.
Durchgeführte Experimente demonstrieren eine interaktive Rekonstruktion von Innenräumen mit einer handgeführten RGBD-Kamera, einer Kinect. Zudem weist das System im Gegensatz zu vergleichbaren hybriden Systemen eine sechsfach höhere Rekonstruktionsrate auf. Bei der Gegenüberstellung anhand eines Benchmark-Datensatzes für RGBD-Sensoren konnte des Weiteren in texturarmen Umgebungen eine Überlegenheit gegenüber punktbasierten Verfahren nachgewiesen werden.

Time: 12.07.2016, 14:00
Location: Room 011 in the Fraunhofer IGD, Fraunhoferstrasse 5, S3|05
Speaker: PeterSheldick (Advisor: Harald Wuest)
Title: "CAD-Model Tracking using RGB-D Cameras" (Master thesis)
Abstract: This thesis deals with the determination of the six DOF of an RGB-D camera relative to a known CAD-Model. Extracting features in image based tracking with no other input data reduces the achievable precision of tracking. This thesis presents methods that use the whole input frame from a depth camera - these are so called "dense" methods. Methods such as ICP, that is used in KinectFusion, and depth image warping, which is used in DVO-SLAM, are compared for the task of CAD-Model tracking. Rendering is used for tracking and both GPU implementations such as OpenGL and CPU ray casting is used to track real depth data.

GRIS Kolloquium

Time: 20.06.2016, 15:30
Location: Room 048 in the Fraunhofer IGD, Fraunhoferstrasse 5, S3|05
Speaker: Michael Behrisch
Doctoral researcher at the University of Konstanz, Germany
Title: "Visual Analytics for the Analysis of Patterns in Matrix Data"
Abstract: Relational data is omnipresent in our computerized society and has found its way into our everyday life. With the growing amounts of relational data the need for analysis techniques dealing with those data sets increases likewise. Typical tasks include not only to visualize the often large and dense data, but also to help the analyst to understand relationships if the data set is multivariate or dynamic in nature.
Several well-known visualization techniques for relational data exist. Next to node-link diagrams, matrix-based representations are another means to visualize and analyze relational data. This compact representation reaches its technical scalability limit not until all display pixels are occupied.
In this talk we will present novel visual interactive techniques, algorithmic approaches and integrated visual analytics systems to support users in navigating and exploring large amounts of relational data. One central research objective is, amongst others, to automatically assess the interestingness of matrix views and show only potentially important matrices from a large exploration space to reduce the users' cognitive overload.
Bio: Michael Behrisch is a doctoral researcher at the University of Konstanz, Germany, since December 2011. He received his MSc. in Information Engineering/Computer Science at the University of Konstanz in 2011 and his BSc. at the Technische Universität Darmstadt. His research interest include the visualization of relational data, pattern analysis in visualizations, and user-centric exploration approaches for large view spaces.

Thesis presentations

Time: 28.04.2016, 14:00 - 14:45
Location: Room 103 in the Fraunhofer IGD, Fraunhoferstrasse 5, S3|05
Speaker: Hannes Kühnel (Advisor: Ralf Gutbell)
Title: "Texturizing and refinement of 3D city models with mobile devices" (Master thesis)

GRIS Kolloquium

Time: 21.04.2016, 13:30
Location: Room 074 in the Fraunhofer IGD, Fraunhoferstrasse 5, S3|05
Speaker: Prof. Dr. Tim Weyrich
Virtual Environments and Computer Graphics, University College London
Visiting Professor at TU Darmstadt
Title: "Application-Driven Appearance Digitisation"
Abstract: The increasingly ubiquitous availability of high-quality digital cameras enables low-cost visual capture and digitisation of real-world objects and phenomena; at the same time, physical output devices, from high-definition screens to computer-controlled manufacturing, are becoming commonplace. This development bears the promise of an even tighter integration of computers into traditional workflows, seamlessly transitioning between the physical and the digital realm. In practice, however, technical off-the-shelf solutions are rarely sufficient to enter previously non-computerised domains.
Tim Weyrich's work focuses on developing novel representations, algorithms and workflows to open up digitisation and modelling of an object's visual appearance for innovative applications. This talk presents such bespoke developments in a number of areas, including special-effects, cosmetics, sculpture and architecture, as well as cultural-heritage preservation, discussing how through careful analysis of traditional problem domains and workflows visual computing can make a difference in previously unexpected ways.

Thesis presentations

Time: 20.04.2016, 10:00 - 10:45
Location: Room 073 in the Fraunhofer IGD, Fraunhoferstrasse 5, S3|05
Speaker: Simon Breitfelder (Advisor: Roman Getto)
Title: "Invertierte GML-Modellierung durch Einpassen und Zusammensetzen von einfachen Strukturen" (Bachelor thesis)
Abstract: In dieser Arbeit wird ein Verfahren zur invertierten Modellierung auf Basis von geometrischen Primitiven vorgestellt. Hierbei wird zunächst die Punktwolkenrepräsentation des Eingabemodells durch mehrere Primitive angenähert. Für die so erzeugten Primitive wird eine zusammenhängende Boundary Representation erzeugt, indem Verbindungsstücke eingefügt werden. Die endgültige Ausgabe des Algorithmus ist ein GML-Programm, welches diese Boundary Representation beschreibt.

Time: 24.03.2016, 10:15 - 11:00
Location: Room 073 in the Fraunhofer IGD, Fraunhoferstrasse 5, S3|05
Speaker: Stepan Konrad (Advisor: Prof. Dr.-Ing. Michael Goesele)
Title: "Reconstruction of Specular Surfaces from Reflectance Correspondences" (Master thesis)
Abstract: Image-based reconstruction of specular surfaces usually requires dense correspondences between image features and points in the environment. In natural environments, these points are usually unknown and correspondences often exist only sparsely between pairs of images. These assumptions complicate the reconstruction problem by introducing many ambiguities which can often only be resolved using regularization of the surface. Only very recently, work has been presented which is able to reconstruct specular surfaces using different kinds of algorithms.
This thesis gives an introduction to the different types of ambiguities and presents a framework which tries to resolve these through regularization using a multi-view approach in combination with a low-parametric surface. The reconstruction method is modeled as an iterative optimization in order to achieve specular consistency. This consistency is based on the laws of reflection applied to the viewing rays which are given by image-to-image features. The framework is capable of processing different kinds of additional input data, e.g. known environmental features or boundary points on the surface.
Synthetic and real-world experiments were executed using both known and unknown feature positions. Results on synthetic datasets show accurate reconstructions even in the presence of specular consistent ambiguities. An adapted outlier removal for feature matching on image series of specular objects was applied to real-wold input data. The results show that it is possible to reconstruct the surface of mirroring objects even with sparse input data.

Time: 24.03.2016, 09:30 - 10:15
Location: Room 073 in the Fraunhofer IGD, Fraunhoferstrasse 5, S3|05
Speaker: Patrick Seemann (Advisor: Prof. Dr.-Ing. Michael Goesele, Dr.-Ing. Simon Fuhrmann)
Title: "Multi-Scale Curvature Field of Triangle Meshes" (Bachelor thesis)
Abstract: We present a novel algorithm that computes a multi-scale curvature field of a triangle mesh. In particular, the algorithm is applicable to meshes produced by image based reconstruction algorithms, which typically comprise geometric features of varying scales. We further show how such a multi-scale curvature field can be helpful for applications like mesh simplification, where the total number of vertices is reduced while preserving surface detail. Our algorithm is based on computing integral invariants using the ball neighborhood and deriving the mean curvature as described by [Yang et al. 2006]. The correct scale for each vertex is automatically determined by the algorithm and does not have to provided as input by the user.

GRIS Kolloquium

Time: 04.03.2016, 11:00
Location: Room 073 in the Fraunhofer IGD, Fraunhoferstrasse 5, S3|05
Speaker: Dr. Hans-Jörg Schulz
Institute for Computer Science, University of Rostock
Institute for Computer Graphics and Vision, Graz University of Technology
Title: "Towards User Support for Expressive Data Visualization"
Abstract: Despite 25 years of research, creating data visualizations remains somewhere between an engineering and a design challenge that is hard to master. Support technologies to aid in this process are still rare, as visualizations are never "drawn once and for all, but constructed and reconstructed until they reveal all the relationships constituted by the interplay of the data." (Bertin 1981) Thus, visualizations are always adapting to different input data and different user interests, which makes it hard to support users in this creative and often chaotic process.
As a first step towards supported data visualization, this talk investigates means that focus on enabling users to produce data visualizations that are expressive – i.e., that correctly reflect the data. To do so, it follows the two principal directions of getting the "right view" for the data and of getting the "right data" for the view. Creating the right view addresses the design challenge of data visualization that is still too often bound by what's available and what's known, instead of what's possible. Whereas, choosing the right data addresses the engineering challenge of data visualization that aims to strike the balance between overloading and oversimplifying the resulting view. In both cases, the shown support techniques put the users in the loop and facilitate their informed choices on data and view as a first step towards more expressive data visualization.

Thesis presentations

Time: 09.02.2016, 10:00
Location: Room 201 in theFraunhofer IGD, Fraunhoferstrasse 5, S3|05
Speaker: Simon Kadel (Advisor: Matthias Borner)
Title: "Automatische und geführte Anreicherung von Building Information Models mit Strukturelementen aus 3D-Punktwolken" (Bachelor thesis)
Abstract: Gebäudeplanung geschieht in heutigen Unternehmen zum großen Teil am Rechner. Dabei kommen sogenannte Building Information Models (BIM) zum Einsatz. Diese enthalten Informationen über das Gebäude wie Grundriss, Fenster, Türen und Verläufe von Leitungen. Wenn ein älteres, bereits bestehendes Gebäude umgebaut, abgerissen oder renoviert werden soll, steht oft kein Modell zur Verfügung. Ein digitales Modell hat den Vorteil, dass die Arbeiten an dem Gebäude effizienter und genauer geplant werden können. Eine Möglichkeit Teile eines BIM zu erhalten, ist das Scannen des Gebäudes und eine anschließende Analyse. Das Verfahren soll möglichst einfach zu bedienen und kostengünstig sein, sowie die Besonderheiten von Low-Cost-3D-Scannern berücksichtigen.

Time: 29.01.2016, 10:00
Location: Room 072 in the Fraunhofer IGD, Fraunhoferstrasse 5, S3|05
Speaker: Felix Hammacher (Advisor: Arjan Kuijper)
Title: "Physical simulation- and reconstruction-framework for shape sensing fabrics" (Master thesis)
Abstract: Over the last decade a large number of prototypes for several research areas in the field of shape sensing have been based on optical tracking devices like the Microsoft Kinect. To overcome the disadvantages of such devices, namely immobility and occlusion of tracked objects, another approach, to which only little attention has been given to so far, is the usage of embedded sensors in fabrics. One of the reasons might be the high effort to manufacture such prototypes with uncertain outcome in terms of matching the requirements of certain use cases. To help developing and planning such devices as well as the used software, a simulation- and reconstruction-framework is introduced in this thesis. Furthermore both parts together enable creating software for a use case even before the hardware is ready. An exemplary workflow, demonstrating how the implemented software can support the development of new applications for shape sensing fabrics, is presented with the Sleeping Posture Recognizer. It uses a blanket with embedded acceleration sensors to determine the sleeping posture of the covered person.

Time: 26.01.2016, 16:30
Location: Room 074 in the Fraunhofer IGD, Fraunhoferstrasse 5, S3|05
Speaker: Pascal Schardt (Advisor: Martin Heß)
Title: "Ghosting- und Poppingdetektor für Image Based Rendering-Sequenzen" (Master thesis)
Abstract: Image Based Rendering-Videosequenzen finden in einer steigenden Anzahl von Bereichen Verwendung, wie in virtuellen Führungen, virtuellen Erkundungen oder TV-Sportübertragungen. Bei der Erstellung solcher Videosequenzen kann es aufgrund verschiedener Ursachen zu Bildartefakten kommen. Diese Artefakte können bisher kaum automatisiert erkannt und deren Störfaktor für menschliche Betrachter ermittelt werden und müssen daher mühsam per Hand gesucht und bewertet werden.
Diese Arbeit beschäftigt sich damit, die am häufigsten vorkommenden und störendsten Artefakte „Popping“ und „Ghosting“ maschinell zu detektieren und die Qualität gefundener Artefakte für einen menschlichen Betrachter anzugeben. Dazu werden unter Berücksichtigung bisheriger, verwandter Arbeiten Algorithmen zur Detektion der beiden genannten Artefaktarten untersucht und weiterentwickelt. Da diese Detektoren keine Ergebnisse der gewünschte Qualität liefern beziehungsweise es für einen Artefakttyp noch keine veröffentlichte Detektionsmöglichkeit gibt, werden neue, eigene Ansätze verfolgt, um Detektionsalgorithmen mit zufriedenstellenderen Resultaten zu implementieren.
Um festzustellen, wie stark die gefundenen Artefakte einem menschlichen Betrachter auffallen, werden für beide Detektoren Qualitätsmetriken aufgestellt, die sich an der menschlichen Wahrnehmung orientieren. Im Zuge der Überprüfung der Güte dieser Qualitätsmetriken wird eine Nutzerstudie durchgeführt, um deren Vergleichbarkeiten mit dem menschlichen visuellen System zu validieren.
Das Ergebnis dieser Arbeit ist ein Detektionsverfahren für Bildartefakte in Image Based Rendering-Videosequenzen, das es erlaubt, solche Sequenzen automatisiert zu verarbeiten. Damit wäre es möglich zum Beispiel eine Aussage zu treffen, ob Teile einer Videosequenz als sehr störend empfunden werden und man versuchen sollte, dieses Video durch erneutes Rendern mit genaueren Tiefenkarten, die durch mehr Eingabebilder erreicht werden können, qualitativ zu verbessern.

GRIS Kolloquium

Time: 26.01.2016, 15:45
Location: Room 074 in the Fraunhofer IGD, Fraunhoferstrasse 5, S3|05
Speaker: Timothy Kol
Computer Graphics and Visualization group
Delft University of Technology
Title: "Geometry and Attribute Compression for Voxel Scenes"
Abstract: Voxel-based approaches are today's standard to encode volume data. Recently, directed acyclic graphs (DAGs) were successfully used for compressing sparse voxel scenes, but they are restricted to a single bit of (geometry) information per voxel. We present a method to compress arbitrary data, such as colors, normals, or reflectance information. By decoupling geometry and voxel data via a novel mapping scheme, we are able to apply the DAG principle to encode the topology, while using a palette-based compression for the voxel attributes, leading to a drastic memory reduction. Our method outperforms existing state-of-the-art techniques and is well-suited for GPU architectures. We achieve real-time performance on commodity hardware for colored scenes with up to 17 hierarchical levels (128K³ voxel resolution), which are stored fully in core.
Bio: Timothy Kol is a PhD candidate at Delft University of Technology in the Computer Graphics and Visualization group. He received his master's degree in Game and Media Technology from Utrecht University in 2014 and his bachelor's degree in Computer Science from Delft University of Technology in 2011. His research interests include alternative representations, real-time rendering and participating media.

Thesis presentations

Time: 26.01.2016, 11:00
Location: Room 103 in the Fraunhofer IGD, Fraunhoferstrasse 5, S3|05
Speaker: Florian Berz (Advisor: Christian Altenhofen)
Title: "Intuitive 3D-Interaktion für Design und Modellierung von volumetrischen Strukturen" (Master thesis)
Abstract: Creating an intuitive way of modeling and designing different structures is a task of high relevance with a wide range of applications in computer graphics. Various options will have to be considered with respect to new developments. For instance, the following options are addressed in this thesis: visualization and interaction. The availability of modern and low-cost 3D hardware provides the users with new possibilities. Visualization and interactions can be performed in 3D and new approaches to applying the interactions are required. In this regard, a combination of zSpace tablet and Leap Motion is proven to fulfill these requirements. Interactions with the mesh can be implemented with the zSpace tablet and Leap Motion is used to control the navigation in scene and menu. Overall, some relevant and new 3D interactions are developed which, amongst other things, allow for the manipulation of inner structures. Those 3D interactions are meaningful, useful and intuitive and therefore, this thesis contributes to the current developments.

Time: 21.01.2016, 14:00
Location: Room 074 in the Fraunhofer IGD, Fraunhoferstrasse 5, S3|05
Speaker: Fabian Schrammel (Advisor: Martin Heß)
Title: "Implementation of a Particle Pusher GPU Port using CUDA" (Bachelor thesis)
Abstract: The Particle-in-cell (PIC) method can be used to simulate the movement of charged particles in an
electromagnetic field. In discrete time steps the particles are moved according to the surrounding fields
on a fixed mesh. The calculation of the particle movement and electric currents are handled by the
so-called Particle Pusher.
An implementation by Grischa Jacobs at the TU Darmstadt is able to compute the particle-mesh
interactions in parallel on MPI coordinated nodes using CPUs and Xeon Phi devices. To reach higher
throughput the Particle Pusher part of the simulation is distributed to as many threads on as many devices
as possible. So far no code to leverage GPUs is included. Using NVIDIA CUDA, a part of the particles
can be simulated using the massively parallel compute power of modern NVIDIA graphics cards. The
concrete goal of this thesis is to determine the basic conditions for this simulation using CUDA. In doing
so, data structure concepts for input and output and the general calculation process are analysed. Also
some basic estimations for memory usage, provided particle counts and execution time are made.

Time: 16.12.2015, 13:45
Location: Room 220 in the Fraunhofer IGD, Fraunhoferstrasse 5, S3|05
Speaker: Joel Müller (Advisor: Christian Altenhofen)
Title: "A Subdivision-Based Approach to the Heat Equation for Simulation-Based Modeling" (Bachelor thesis)
Abstract: In this thesis a subdivision-based method is presented for calculating numerical solutions to differential equations on the basis of a geometric representation that is also well suited for modeling. While modern CAD programs mainly use continuous representations, like B-splines or NURBS, numerical methods like FEM require a discrete mesh to perform the calculation on. The conversion between these two representations can become a hugely timeconsuming process. Utilizing the same representation for modeling and simulating objects speeds up the whole engineering process, as the need for mesh generation is eliminated. This also reduces the error made by approximating the geometry. As subdivision schemes are intuitive and efficient to use for modeling and visualizing complex geometries, they serve well as a basis for this method. The presented method is based on Chaikin's algorithm for one-dimensional objects and utilizes Catmull-Clark surfaces to represent two-dimensional objects. On the basis of these two subdivision schemes, solutions to the heat equation are generated, demonstrating the applicability of the approach. The exactness of this solution and the performance of the algorithm are compared to a traditional FEM approach to the heat equation.

GRIS Kolloquium

Time: 15.12.2015, 16:30
Location: Room 074 in the Fraunhofer IGD, Fraunhoferstrasse 5, S3|05
Speaker: Prof. Dr. Alexei Sourin,
Nanyang Technological University (NTU) &
Fraunhofer Project Centre for Interactive Digital Media at NTU
Title: "Visual Immersive Mathematics: From Ancient Greeks to Digital Natives"
Abstract: In this 45 min talk, Dr. Sourin will first look through the history of a few major events that changed "geometric" mentality of people. He will then highlight cognition problems arising when working in computer graphics and virtual reality. Finally, function-based approach to shape modelling and visualization will be discussed including FRep representation, function-based extension of VRML and X3D, as well as tangible images and videos.
Bio: Dr. Alexei Sourin is a university professor at Nanyang Technological University (NTU) in Singapore. He is also holding a concurrent appointment of Deputy Director at Fraunhofer IDM@NTU-The Fraunhofer Project Centre for Interactive Digital Media at NTU. He is a co-founder of the Function Representation concept (known in computer graphics as FRep) which he published in nearly 200 papers and books on shape modeling, web visualization, and virtual reality. Dr. Sourin is a coordinator of two joint PhD degree programmes between Nanyang Technological University in Singapore, Technischen Universität Darmstadt, (Germany) and Graz University of Technology (Austria). Dr. Sourin is a recipient of many scientific awards. He is a Chair of the workgroup "Computer Graphics and Virtual Worlds" of the International Federation for Information Processing (IFIP). Dr. Sourin is a Senior Member of IEEE and a member of ACM SIGGRAPH. He is on the Editorial Boards as well as a Guest Editor of several international journals. He was General and Program Chair of 12 conferences as well as member of program committees of over 120 international conferences. He is a coordinator of the International Conferences on Cyberworlds. More information can be found at

Thesis presentations

Time: 14.12.2015, 09:00
Location: Room 242 in the Fraunhofer IGD, Fraunhoferstrasse 5, S3|05
Speaker: Amir Kalali Emghani (Advisor: Florian Jung)
Title: "Vertebrae detection in head and neck MRI images using a particle filter method based on image features" (Master thesis)
Abstract: Detection of the human spine in MRI scans is a frequently used task in medical image analysis for the diagnosis of spinal conditions and diseases. MRI scans provide important information about relevant soft tissue. However, manual detection of vertebrae is an elaborate and error-prone process. This work aims at the development and evaluation of a semi-automatic method for vertebrae detection in head and neck MRI images. The goal is to correctly position an available articulated atlas that consists of statistical shape models of bones.
The presented approach is based on a probabilistic graphical model for modelling the structure of the vertebral column. A particle filter is used to define the position of vertebral bodies in relation to previously detected vertebral bodies. Two models are used to model the geometric constraints of the position, the size and the orientation of the vertebral bodies on the one hand and to describe the vertebral body appearance with extracted image features on the other hand. The advantage of this approach is the probabilistic analysis of potential vertebral bodies: It does neither depend on exact prior knowledge about the anatomical structure of the vertebral column nor any training data to learn the image features is needed.
The presented approach is tested and the vertebrae detection results are evaluated on 22 MRI scans. The MRI scans are T1-weighted sagittal 2D slices. Subsequently the method is used to initially position the articulated atlas in the MRI images for further processing. Visual inspections show a noticeable improvement of the initial position of the atlas compared to the approach used so far.

Time: 11.12.2015, 10:00
Location: Room 242 in the Fraunhofer IGD, Fraunhoferstrasse 5, S3|05
Speaker: Joachim Loge (Advisor: Andreas Braun)
Title: "Evaluating capacitive distance and haptic input modalities for sound synthesis combined with Augmented Reality" (Master thesis)
Abstract: In den letzten Jahren haben Geräte wie die Microsoft Kinect oder die Leap Motion berührungslose Gestenverfolgung erschwinglich gemacht.
Während dies eine sehr schnelle und natürliche Form der Eingabe ist, können viele Anwendungen von einer Kombination der Freiraum-Interaktion mit Berührungserkennung profitieren. Der CapTap ist ein interaktiver Tisch, der kapazitive Entfernungssensoren und akustische Berührungserkennung kombiniert und damit viele Interaktionsmodi ermöglicht. In dieser Arbeit wurden diese Fähigkeiten in dem Kontext musikalischer Kontrollszenarien evaluiert. Das Ziel der Forschung war, die Komplexität der Interaktionsmodi zu bestimmen und Ergebnisse darüber zu sammeln, wie ansprechend und motivierend die Testpersonen die musikalische Interaktion finden. Ein weiterer Fokus lag auf dem Vergleich einer Visualisierung mit erweiterter Realität und einem gewöhnlichen Bildschirm, um Daten über die bevorzugte Visualisierungstechnologie zu erhalten. Außerdem wurde die akustische Berührungserkennung des CapTap um eine Eingabe mit dem Fingernagel erweitert. Die technische Evaluation befasste sich mit der Zuverlässigkeit von drei Berührungseingaben und ergab eine durchschnittliche Erkennungsrate von 85%.
Die Benutzerevaluation ergab, dass alle Modi hoch stimulierend sind und das Bedürfnis unterstützen, sich weiterzuentwickeln. Während die ein- und zweihändigen berührungslosen Interaktionsmodi ohne viel Übung eingesetzt werden konnten und als sehr attraktiv bewertet wurden, erhielten die kombinierten Berührungs- und Freiraum-Interaktionsmodi schlechtere Bewertungen. Der Vergleich zwischen einer Visualisierung mit erweiterter Realität und einem gewöhnlichen Computer-Monitor ergab eine klare Präferenz für den gewöhnlichen Bildschirm.

Time: 20.11.2015, 11:00
Location: Room 220 in the Fraunhofer IGD, Fraunhoferstrasse 5, S3|05
Speaker: Marko Rücker (Advisor: Matthias Borner)
Title: "Schnelle und wasserdichte Triangulierung von Punktewolken für die Strömungssimulation" (Bachelor thesis)
Abstract: In der frühen Designphase eines realen Objektes (z.B. mittels Tonmodell) ist eine schnelle Beurteilung darüber notwendig, ob das erstellte Objekt vorher definierten Anforderungen gerecht wird. Zum Beispiel soll frühzeitig und mit geringem Aufwand herausgefunden werden, ob das Objekt gewisse Strömungseigenschaften oder statische Eigenschaften besitzt. Traditionell wird ein reales Objekt nach dem Modellieren in einen Windtunnel gegeben, um Strömungseigenschaften festzustellen. Oder das Objekt wird aufwändig digitalisiert und Simulationen auf dem virtuellen Modell gerechnet. Beide Varianten können für eine einzige Iteration viele Stunden bis hin zu einigen Arbeitstagen in Anspruch nehmen. Für einen Zeit-effizienten Formfindungsprozess ist es wünschenswert, dass der Workflow von Modellieren über Scannen bis hin zum Simulieren so kurz wie möglich gestaltet wird. Daraus folgt die Notwendigkeit einer schnellen Digitalisierung des Objektes, welche in einer wasserdichten Triangulierung resultiert, auf der eine Simulation (z.B. von Strömungen) korrekt durchgeführt werden kann.

Time: 20.11.2015, 09:00
Location: Room 242 in the Fraunhofer IGD, Fraunhoferstrasse 5, S3|05
Speaker: Lukas Appelhans (Advisor: Florian Jung)
Title: "Adaptive voxel-based classifier for semi-automatic segmentation of tumors in the head and neck area based on T2-weighted MRI image data" (Bachelor thesis)
Abstract: Measuring the size and location of a tumor is a major part of cancer staging and thus also crucial to
plan treatment and predict the success chances of the same. Both properties can be extracted from a
We present a new method for semi-automatic segmentation of tumors in the head and neck area using
MR images. The new method incorporates known segmentations that were manually created by medical
doctors. Other than that the only user interaction needed is setting a seed point.
After the seed point and an input image are entered, the algorithm starts by searching for a similar
one in the database. The underlying assumption is that the intensities of a tumor in two comparable
images also have comparable values. Using histograms for both the database image as well as the
manual segmentation of it, the intensities that are likely to be featured in the tumor are calculated. After
creating a basic segmentation, the actual tumor is extracted using opening, closing and a connected
threshold filter.
The algorithm was developed using five datasets of T2-weighted MR images with a leave-one-out cross
validation technique. When comparing the generated tumor segmentations with the manual ones, they
had a DSC in the range of 0.41 and 0.77, with an average of 0.60. Furthermore the new method was
also tested on lymph nodes. Further suggestions for improvements are given.

Time: 16.11.2015, 16:00
Location: Room 048 in the Fraunhofer IGD, Fraunhoferstrasse 5, S3|05
Speaker: Barbara Zöller (Advisor: Tatiana von Landesberger)
Title: "Visuelle Analyse von stochastischen Verspätungen im Zugverkehr" (Master thesis)
Abstract: Zur Planung von Zugfahrten werden im Normalfall Verbindungsplaner, wie zum Beispiel die Reiseauskunft der Deutschen Bahn AG verwendet um Informationen über mögliche Verbindungen hinsichtlich Daten wie Abfahrts- und Ankunftszeit sowie Reisedauer zu erhalten. Allerdings ist für den Reisenden genauso das Risiko von Verspätungen bei der Auswahl der geeignetsten Verbindung relevant. Diese wird aber bisher nicht in Planer dargestellt. Dabei wäre mehr Planungssicherheit wünschenswert, um nicht zum Beispiel zu einem wichtigen Termin eine Stunden früher anreisen zu müssen, aus Angst vor möglichen Verspätungen. Daher ist die Idee, Verbindungsplaner zu erstellen, die Informationen zu Verspätungen enthalten um damit die Fahrtplanung zusätzlich zu unterstützen. So könnten dadurch für Verspätung anfällige Züge vermieden oder Umstiege mit geringen Erfolgsaussichten umgangen werden und so in Konsequenz die Fahrt für den Reisenden deutlich angenehmer gemacht werden. In dieser Masterarbeit werden in einem ersten Schritt zwei verschiedenen Visualisierungsformen für Verspätungen im Zugverkehr entwickelt. Im nächsten Schritt werden diese beiden mit der Reiseauskunft der Deutschen Bahn AG als Referenz in einer Online-Umfrage verglichen. Hierbei werden auch die persönlichen Präferenzen des Bahnfahrers berücksichtigt, mit der Absicht diese in zukünftige Visualisierungen mit einfließen zu lassen um den Entscheidungsprozess für den Bahnfahrer weiter zu vereinfachen. Schließlich erfolgen in einem letzten Schritt die Analyse der Umfrageergebnisse und Ratschläge für die Weiterentwicklung der Visualisierungen.

Time: 11.11.2015, 14:00
Location: Room 048 in the Fraunhofer IGD, Fraunhoferstrasse 5, S3|05
Speaker: Kamran Yaqub (Advisor: Dirk Burkhardt)
Title: "Distributed Social Media Analysis on Microblog-Services for Policy Modeling" (Master thesis)

GRIS Kolloquium

Time: 09.11.2015, 10:30
Location: Room 073 in the Fraunhofer IGD, Fraunhoferstrasse 5, S3|05
Speaker: Kyle Rector
Department of Computer Science and Engineering
University of Washington
Title: "Design and Evaluation of Eyes-Free Exercise Technologies"
Abstract: People who are blind or low-vision may have a harder time participating in exercise classes due to inaccessibility, travel difficulties, or lack of experience. In this talk, I will present my dissertation research where I discuss the invention of Eyes-Free Yoga, a yoga exergame accessible to people who are blind and low-vision by providing auditory instructions and feedback. I will present the results from both a lab study with 16 participants and an 8-week deployment study with 4 participants, all of whom are visually impaired. Finally, I will describe a project with in-depth interviews and surveys that inform the design of future eyes-free exercise technologies. My contributions are the design, development, and evaluation of technologies that will benefit blind and low-vision wellness.
Bio: Kyle Rector is a sixth year PhD candidate in the Department of Computer Science and Engineering at the University of Washington, co-advised by Julie Kientz and Richard Ladner. She is interested in developing Eyes-Free Exercise Technologies, specifically exercise games for people who are blind or low-vision. She is a Google PhD Fellow (2015), and previously was an NSF Graduate Research Fellow (2012-2015), Google Anita Borg Scholar (2010), and Palantir Scholarship for Women in Technology Semi-Finalist (2013).
Her research has been recently covered by MIT Technology Review, Microsoft, Gizmag, GeekWire, and c|net. Kyle received her MS from the University of Washington (2012), and her BS from Electrical & Computer Engineering and Computer Science from Oregon State University (2010).

Thesis presentations

Time: 26.10.2015, 16:30
Location: Room 074 in the Fraunhofer IGD, Fraunhoferstrasse 5, S3|05
Speaker: Felix Gorschlüter (Advisor: Nicolas Weber)
Title: "Eine quantitative Evaluation von Farbtransferverfahren für Augmented Reality" (Master thesis)

GRIS Kolloquium

Time: 16.10.2015, 11:00
Location: Room 074 in the Fraunhofer IGD, Fraunhoferstrasse 5, S3|05
Speaker: Prof. Dr. Sai-Kit Yeung
Vision, Graphics and Computational Design (VGD) Group
Singapore University of Technology and Design (SUTD)
Title: "Data-driven Computer Graphics Modeling"
Abstract: Modeling is a key step in the computer graphics pipeline. Computer animations, games, computer-aided design and manufacturing create a strong need for high quality 3D models, yet the manual creation of 3D models is a very tedious process which involves a lot of time and human effort.
In this talk, we explore how the rapidly increasing quantity of user-created 3D data and real-world scene data publicly available can be adapted to facilitate the essential task of modeling. We first discuss the conceptual innovations inherent to model synthesis through data driven optimization, along with the advantages of and considerations in its application. We then tackle various challenging modeling problems within our novel framework. We will present data-driven optimization methods for virtual world modeling, virtual character modeling and zoomorphic design. We will conclude the talk by discussing future directions in modeling and some ongoing projects in our group.
Sai-Kit Yeung is an Assistant Professor at the Singapore University of Technology and Design (SUTD), where he leads the Vision, Graphics and Computational Design (VGD) Group. Before joining SUTD, he had been a Postdoctoral Scholar in the Department of Mathematics, University of California, Los Angeles (UCLA). He was also a visiting student at the Image Processing Research Group at UCLA in 2008 and at the Image Sciences Institute, University Medical Center Utrecht, the Netherlands in 2007. He received his PhD in Electronic and Computer Engineering from the Hong Kong University of Science and Technology (HKUST) in 2009. He also received a BEng degree (First Class Honors) in Computer Engineering in 2003 and a MPhil degree in Bioengineering in 2005 from HKUST. His research interests include computer vision, computer graphics and computational fabrication.

Thesis presentations

Time: 15.10.2015, 10:00
Location: Room 048 in the Fraunhofer IGD, Fraunhoferstrasse 5, S3|05
Speaker: Sachin Pattan (Advisor: Dirk Burkhardt)
Title: "Distributed Search Intention Analysis for User-Centered Visualizations" (Master thesis)
Abstract: In recent years, Web Search Engines (WSEs) are the most used Information Retrieval systems around the world. As the information available increases explosively, it becomes more difficult to fetch the information meeting the preferences. This calls for a deep study of knowledge about the users' pre-knowledge and intentions and hence is a critical area of research in many organizations. There are many existing implementations of search intention analysis in some famous search engines such as Google, Bing etc. There are also several researches which propose few approaches for search intention analysis. Still there is a lack of techniques for intention mining in the field of semantic visualizations as they are designed to provide the visual adaptations especially for Exploratory queries. Hence, it is critical to identify the exploratory and targeted search queries. Also, the advancements in the technologies of distributed software systems make them to be applicable in all the systems which need to do some sort of distribution of load. The search intention analysis requires the distribution of load to do parallel processing of intention mining.
In this thesis, a new approach for classifying the search intention of users' in a distributed set up is described along with a deep study on existing approaches with their comparison. The approach uses the efficient parameters: word frequency, query length and entity matching for differentiating the user query into exploratory, targeted and analysis search queries. As the approach focuses mainly on frequency analysis of the words, the same is done with the help of many sources of information such as Wortschatz frequency service by university of Leipzig and the Microsoft Ngram service. The model is evaluated with the help of a survey tool and few Machine Learning techniques. The survey was conducted with more than hundred users and on evaluating the model with the collected data, the results look satisfactory.

GRIS Kolloquium

Time: 05.10.2015, 10:00
Location: Room 074 in the Fraunhofer IGD, Fraunhoferstrasse 5, S3|05
Speaker: Dr. Kenneth Vanhoey
GraphDeco group, INRIA Sophia-Antipolis Méditerrannée
Title: "From real data to virtual copies: a few contributions"
Abstract: This talk will cover two applications related to digitization of the real world: one concerning cultural heritage and another on example-based texture synthesis.
I) Treatment of 3D models with acquired radiance.
Vision and computer graphics communities have built methods for digitizing, processing and rendering 3D objects. There is an increasing demand coming from cultural communities for these technologies, especially for archiving, remote studying and restoring cultural artefacts like statues, buildings or caves. Besides digitizing geometry, there can be a demand for recovering the photometry with more or less
complexity: simple textures (2D), light fields (4D), SV-BRDF (6D), etc.
This talk will cover several contributions that can be situated in the pipeline for constructing and treating surface light fields represented by hemispherical radiance functions attached to a surface. First, we will tackle robustness issues in the aspect reconstruction from photographic data resulting from real-world on-site acquisitions [VSLD13]. And secondly we will present a simplification technique for this data, which locally minimizes the loss of both geometric and photometric detail [VSKLD15].
II) Example-based real-time texturing.
Generating textures in real-time is a must in videogames or other real-time animated worlds. But to create textures quickly, they need to be generated at runtime from compact data stored on the GPU. In this talk we will explain why, and overview two contributions that improve over the state of the art in on-the-fly texture synthesis. The first [VSGLD13] is a new tiling algorithm that augments the amount of variety that can be generated with a single repetitive tile. This is achieved by stochastic exchanges of part of its content in a seamless way. Above that, a multi-scale transition mechanism is added to generate visual detail. The second [GSVDG14] will propose a new noise model for procedural texturing, which are textures stored as compact functions.
This new function has been built so that it can be tuned to reproduce an input exemplar containing structured elements in a procedural way, which was unseen before.
[VSLD13] Robust Fitting on Poorly Sampled Data for Surface Light Field Rendering and Image Relighting ; Computer Graphics Forum, vol. 32, issue 6.
[VSGLD13] On-the-Fly Multi-Scale Infinite Texturing from Example ; Proceedings of "ACM SIGGRAPH Asia 2013" conference ; Transactions on Graphics, vol. 31, issue 6.
[GSVDG14] Local random-phase noise for procedural texturing ; Proceedings of "ACM SIGGRAPH Asia 2014" conference ; Transactions on Graphics, vol. 32, issue 6.
[VSKLD15] Simplification of Meshes with Digitized Radiance ; Proceedings of the "Computer Graphics International 2015" conference ; The Visual Journal, vol. 21, issue 6-8.
Kenneth Vanhoey is a postdoctoral fellow at Inria, France, working in the GraphDeco group with George Drettakis. He obtained a B.Sc. in 2008, a M.Sc. in 2010 and a Ph.D. in Computer Science in 2014 at the University of Strasbourg, France. He worked there in the IGG group of the ICube laboratory with Basile Sauvage and Jean-Michel Dischler. His Ph.D. thesis focuses on reconstructing, visualizing and post-treating the appearance of 3D models in the scope of 3D acquisition of real objects. His current research interests lie with texturing and image editing. More information on his research can be found at

Thesis presentations

Time: 02.10.2015, 13:00
Location: Room 048 in the Fraunhofer IGD, Fraunhoferstrasse 5, S3|05
Speaker: Fabio Arnold (Advisor: Martin Knuth)
Title: "Approximation von Reflexionsmodellen für das interaktive Kleidungsdesign unter natürlicher Beleuchtung" (Bachelor thesis)

Talk and discussion

Time: 25.09.2015, 14:00
Location: Room 074 in the Fraunhofer IGD, Fraunhoferstrasse 5, S3|05
Speaker: Prof. Donald Lupo
Department of Electronics and Communications Engineering Laboratory for Future Electronics
Tampere University of Technology
Title: "Making the Internet of Everything Environmentally Sustainable, and Printable: Printed Energy Harvesting, Storage and Circuits for Distributed Smart Objects"
Abstract: There is a lot of talk about putting electronic sensors "everywhere", enabled both by miniaturization of classic Si electronics and advances in printed electronics, and the "Internet of Things" is a hot topic.
However, sensors everywhere require power everywhere, and hundreds of billions of small objects fitted with batteries would be a waste disposal nightmare. An alternative is the harvesting of ambient energy, e.g. from light, RF radiation and movement, combined with interim storage in completely non-toxic printed supercapacitors. A further issue is silicon manufacturing capacity, which is nowhere near large enough to provide hundreds of billions of CMOS chips in the long term. At Laboratory for Future Electronics a large part of our research is aimed at making the Internet of Everyting sustainable, printable, and flexible. In this talk I will introduce our research on a hopefully not too technical level and introduce in particular two major project that are specifically aimed at this goal: "Printed, energy Autonomous UniversaL platform for multifunctional wireless sensors and devices (PAUL)", and "THE NAKED APPROACH: Nordic perspective to gadget-free hyperconnected environments", a cooperation with several Finnish organizations.
Bio: Donald Lupo joined the Department of Electronics and Communications Engineering at Tampere University of Technology as professor for electronic materials and the Head of Laboratory for Future Electronics (LFE) in August 2010 after a diverse career in industrial research and development in functional materials for photonics and electronics. He obtained his Ph.D. in physical chemistry at Indiana University-Bloomington, USA in 1984 and spent the next 24 years working in chemical, electronic and display industries, and as an independent consultant, working for and with companies such as Hoechst AG, Sony Europe, NTera, Samsung, UPM Kymmene and Merck. During his industrial career he led groundbreaking work in organic nonlinear optics, polymer LEDs, solid state dye solar cells and paper-like displays. He is author on over 60 publications and inventor on over 40 patents and applications.
He serves as an external expert in the OLED and printed electronics fields for the European Commission, is an active member of the Organic Electronics Association roadmap team and has served on the technical advisory boards of Thin Film Electronics AB, Nano Eprint Ltd., Centre for Process Innovation (CPI) and the EPSRC Centre for Innovative Manufacturing in Large Area Electronics (CIMLAE).

Thesis presentations

Time: 03.08.2015, 13:00
Location: Room 073 in the Fraunhofer IGD, Fraunhoferstrasse 5, S3|05
Speaker: Reimond Retz (Advisor: Kawa Nazemi)
Title: "Assisted Visual Data Exploration for Discovering Information in Digital Libraries" (Master thesis)

Time: 24.07.2015, 14:00
Location: Room 072 in the Fraunhofer IGD, Fraunhoferstrasse 5, S3|05
Speaker: Daniel Herb (Advisor: Reimar Tausch)
Title: "Entwicklung und Evaluation eines echtzeitfähigen Streifenlichtscanners für schnelle und automatisierte 3D-Digitalisierung" (Master thesis)
Abstract: Diese Arbeit befasst sich mit der Entwicklung und Evaluation eines echtzeitfähigen Streifenlichtscanners zur schnellen und automatisierten 3D-Digitalisierung. Der konkrete Anwendungsfall besteht in der massenhaften Digitalisierung von Kulturgütern zur Archivierung und Forschungskollaboration. Als Grundlage dient das Flying Triangulation Prinzip, bei dem ein Sensor handgeführt zur Rekonstruktion verwendet werden kann und diese dem Nutzer in Echtzeit zur Verfügung steht. Ein Ziel dieser Arbeit ist die Aufarbeitung, Klarstellung und Verbesserung bestimmter Komponenten in dem Verfahren, so dass ein robuster und automatisierter Einsatz an einem bionisch inspirierten, nachgiebigen Roboterarm möglich ist. So kann durch eine neu entwickelte Modellfunktion die Systemkalibrierung hinsichtlich der zu findenden Parameter deutlich verbessert werden, wodurch aus Nutzersicht eine einfachere Kalibrierung möglich ist. Außerdem wird für die Registrierung ein Verfahren zur globalen Optimierung verwendet, welches bereits während der Datenakquise ausgeführt werden kann, wodurch der Nachverarbeitungsschritt entfällt. Neben dem eigentlichen Verfahren konnte außerdem eine Simulationsumgebung zur Generierung synthetischer Bilddaten sowie ein realer Hardwareaufbau entwickelt werden. In einem Evaluationskapitel werden für die synthetisch generierten Daten unterschiedliche Parameter variiert, um so deren Auswirkungen auf das Rekonstruktionsergebnis zu untersuchen. Abschließend wird mit einem exemplarischen Objekt ein Praxistest mit dem realen Aufbau durchgeführt.

Time: 09.07.2015, 14:00
Location: Room 073 in the Fraunhofer IGD, Fraunhoferstrasse 5, S3|05
Speaker: Julian von Wilmsdorff (Advisor: Arjan Kuijper)
Title: "Activity Recognition based on Electric Potential Sensing" (Master thesis)
Abstract: Electric fields are influenced by the human body and other conducting materials. This property can be used to detect presence of human bodies. For the detection of presence and activity recognition, mostly capacitive measurement techniques are used. Despite the fact that capacitive sensing is an fairly old technology, since it has been around since the 1920's, it is still a hot topic of ongoing research works. Today, capacitive measurement techniques are used in touch-screens, in the automobile industry and many other fields of Ubiquitous Computing.
But a drawback of the capacitive technology is the energy consumption, which is an important aspect of mobile devices. That is why, in this thesis, i investigate the potential of electric potential sensing (EPS), a purely passive capacitive measurement technique, which can be implemented with an extremely low power consumption.
First, the most commonly used capacitive measurement techniques will be analyzed and how they work. This is done to understand the pros and cons of electric potential sensing compared to other technologies. After analyzing electric potential sensing and related capacitive measurement techniques, we will have a closer look at some possible areas of application of electric potential sensing in an explorative study. Hence, multiple experiments, involving electric potential sensing in various environmental settings for different use-case scenarios, will be conducted. This is done to evaluate the best use-case for this technology. Then, after selecting the most suitable use-case for activity recognition with EPS, two sensor systems are developed, discussed and evaluated. At the end, the benefits and limitations of EPS will be concluded with regards to capacitive sensing.

Time: 17.06.2015, 09:00
Location: Room 324 in the Fraunhofer IGD, Fraunhoferstrasse 5, S3|05
Speaker: Lisa Niel (Advisor: Georgios Sakas)
Title: "Semi-Automatic Landmark Extraction for Intra-Operative Brain Shift Correction" (Master thesis)
Abstract: Im Rahmen der Masterarbeit wurden verschiedene Landmarken im Gehirn auf ihre Eignung zur Segmentierung in MRT und Ultraschall Aufnahmen und zukünftiger Registrierung untersucht. Basierend darauf wurde eine robuste Segmentierung der Ventrikel in MRT- Aufnahmen implementiert.

Time: 16.06.2015, 14:00
Location: Room 103 in the Fraunhofer IGD, Fraunhoferstrasse 5, S3|05
Speaker: Sven Frese (Advisor: Michael Kraemer)
Title: "Secure Cloud-Based Risk Assessment for Urban Areas" (Master thesis)
Abstract: Während die Zahl der Menschen, die in urbanen Zentren leben, deutlich zugenommen hat, ist auch die Anzahl der terroristischen Ereignisse in den letzten Jahrzehnten weltweit stark angestiegen. Besonders dicht besiedelte Gebiete sind anfällige Ziele, weshalb Stadtplaner Sicherheitsmaßnahmen im Planungsprozess besonders berücksichtigen und evaluieren müssen. Um Stadtplaner in diesem Prozess zu unterstützen, stellen wir ein System-Design für eine Web-Anwendung vor, mit der Stadtplaner eine Risikoanalyse für Stadtgebiete auf Basis digitaler Stadtmodelle durchführen und sich die Ergebnisse im Browser visualisieren lassen können. Neben einem Web-Portal beinhaltet unser System einen Web-Service auf Basis einer Mircoservices-Architektur, der Komponenten des CityServer3D, welcher am Fraunhofer IGD entwickelt wird, nutzt, um räumliche Daten zu verarbeiten. Die Risikoanalyse wird mit Hilfe des am Fraunhofer EMI entwickelten VITRUV Tools durchgeführt. Wir verwenden Ciphertext-Policy Attribute-Based Encryption (CP-ABE) in Kombination mit symmetrischer Verschlüsselung, um Nutzerdaten mit eingebetteten Zugriffsrichtlinien auf skalierbarem Cloud-Speicher zu sichern, ohne dass Unbefugte die Daten einsehen können. Wir verschlüsseln Koordinaten von Gebäuden mit Order-Preserving Encryption (OPE), um Abfragen zu Gebäudedetails auf Basis der verschlüsselten Daten zu ermöglichen. Unsere Ergebnisse zeigen, dass unser System-Design skalierbar im Bezug auf die Datenverarbeitung und Verschlüsselung ist.
Allerdings erzeugt die Verschlüsselung mittels CP-ABE einen erheblichen Overhead und ist daher nicht für die Einbettung von komplexen Zugriffsrichtlinien in verschlüsselte Daten geeignet, ohne dass die Reaktionszeiten von Web-Anwendungen signifikant beeinträchtigt werden.

Time: 29.05.2015, 14:00
Location: Room 073 in the Fraunhofer IGD, Fraunhoferstrasse 5, S3|05
Speaker: Michael Alekseew (Advisor: Andreas Braun)
Title: "Hochpräzise Algorithmen zur Detektion von Armen für Interaktionssysteme aus kapazitiven Abstandssensoren" (Master thesis)
Abstract: In dieser Master-Thesis werden basierend auf den Sensordaten des kapazitiven Nahbereichssystemes CapTap eine Handmittelpunktpositions-Erkennung implementiert und getestet. Dazu werden die 24 Sensorwerte verwendet und Referenzdaten einer Handerkennung über die Microsoft Kinect Version 2 kombiniert. Diese Werte werden in einer Trainingsmatrix zusammengefasst und auf diesen Trainingsdaten eine Random Decision Forest (RDF) Algorithmus trainiert.
Dieser RDF wird dann verwendet um alleine von den Sensorwerten auf Handmittelpunktpositionen schließen zu können.

Time: 21.05.2015, 13:00 - 14:00
Location: Room 140 in the Fraunhofer IGD, Fraunhoferstrasse 5, S3|05
Speaker: Felix Bernhard (Advisor: Timo Engelke)
Title: "User Friendly Calibration for Tracking of Optical Stereo See-Through Head Worn Displays for Augmented Reality" (Bachelor thesis)
Abstract: In recent time devices like Google Glass and Oculus Rift gained a lot of public attention. So the field of Virtual and Augmented Reality has become a more and more attractive field of study. This thesis focuses on the Epson Moverio BT-200, an Optical Stereo See-Through Head Worn Display (OST-HWD or OST-HMD) which can be used for Augmented Reality. Augmented Reality devices like this have to be calibrated. This means, that one has to find a configuration, that aligns the image shown on the displays with the environment, which is observed by the built-in camera. If this is not done, the augmented virtual image would not align with the real world. The process of this calibration approach is divided into two stages, like Owen et al. [24] proposed in earlier work, but with less constrains for the positions of the cameras, which makes it easier to use. The first stage is the hardware calibration. It defines the camera intrinsics, containing focal length and principal point, the camera extrinsics, representing the position of the device in world coordinates, as well as the position of the virtual plane onto which the virtual image is projected. The second stage is the user calibration, which calibrates the eye positions for every individual user. An earlier approach by the Fraunhofer IGD [31][32] has been adopted in this work, aiming at a more user friendly suite for the calibration of OST-HWD devices. Therefor both of the aforementioned stages are combined in a new quick step-by-step installation wizard, which is written in HTML and JavaScript to ensure easy usability. Furthermore the VisionLib, developed by Fraunhofer IGD, is used to for image detection and processing [7]. It has been extended by a new minimization model in order to simplify and robustify the calculations of the virtual plane. In addition to that the required hardware components, including camera and calibration rig, were simplified. The implemented software has been evaluated for its results of the computed virtual plane, intrinsic data and eye positions of the user. Furthermore a user study was conducted to rate the usability of the calibration process.

Time: 08.05.2015, 10:00
Location: Room 073 in the Fraunhofer IGD, Fraunhoferstrasse 5, S3|05
Speaker: Andreas Muttscheller (Advisor: Martin Heß)
Title: "Augmented Reality in der Lagerlogistik" (Bachelor thesis)
Abstract: Die Bachelorarbeit wird das Thema Augmented Reality (AR) in der Lagerlogistik behandeln, genauer die Wegfindung eines Lagermitarbeiters zu einem Punkt im Lager.
Zunächst wird die Positionsermittlung innerhalb des Lagers betrachtet. Wie kann man möglichst genau ein Gerät orten und welche Technologien sind dafür notwendig. Dafür werden unter anderem WLAN, Beacons, QR/AR Codes auf ihre Tauglichkeit untersucht. Auch der Blickwinkel sowie die Laufrichtungserfassung muss untersucht werden. Verschiedene Techniken, wie z.B. der Kompass in einem Gerät, werden daraufhin analysiert. Des Weiteren wird geprüft, wie sich die räumlichen Gegebenheiten in einem Lager auf die Funktechnologien (WLAN, Bluetooth) auswirken. Da die Lager und Regale größtenteils aus Metall gefertigt sind, muss deren Beeinflussung auf das Signal untersucht werden. Es werden außerdem verschiedene mobile AR SDKs untersucht, mit Hilfe deren die Informationen dem Nutzer angezeigt werden. Um das Lagerlayout und den Laufweg zu ermitteln wird die Navigation mit dem Lagerführungssystem LFS verbunden. Dort lässt sich bereits das Lagerlayout pflegen und auch die Wegberechnung anhand verschiedener Laufwegstrategien ausführen. Dieses Modul nennt sich "Transport Leit System" TLS und wird für die Wegberechnung genutzt.
Für einen Prototyp der Navigationslösung wird der folgende Anwendungsfall betrachtet: Ein Lagermitarbeiter steht an einem beliebigen Punkt im Lager und soll an einen anderen Punkt geführt werden. Dieser Punkt wird vom LFS vorgegeben und ist in diesem Fall die nächste Pickposition bei der Kommissionierung. Die Anwendung ermittelt dabei die Position des Mitarbeiters und lässt den Laufweg über das TLS berechnen. Dieser wird dem Nutzer anschließend auf seinem Gerät angezeigt und aktualisiert, sobald er sich bewegt.
Zum Ende wird zusammengefasst, ob die angestrebte Lösung produktiv einsetzbar ist.

Time: 29.04.2015, 16:00
Location: Room 324 in the Fraunhofer IGD, Fraunhoferstrasse 5, S3|05
Speaker: Olga Koziol (Advisor: Felix Brodkorb)
Title: "Edge Bundling and collision-free Routing" (Practical course)

Time: 16.04.2015, 09:00 - 10:00
Location: Room 324 in the Fraunhofer IGD, Fraunhoferstrasse 5, S3|05
Speaker: Hermann Reichert (Advisor: Georgios Sakas)
Title: "Kalibrierung von räumlich verfolgten Ultraschall-Sonden" (Bachelor thesis)
Abstract: Das Wissen über die Position von 2D-Ultraschallbildern im 3D-Raum ist eine wertvolle Information, die genutzt wird, um Volumen zu rekonstruieren. Diese Positionsbestimmung erfolgt über ein Trackingsystem, welches die Position der Ultraschallsonde verfolgt. Die Beziehung zwischen der Position der Sonde und der Position des eigentlichen Ultraschallbildes wird durch eine Transformationsmatrix beschrieben, die mittels einer Kalibrierung bestimmt werden muss.
Diese Arbeit gibt einen Überblick der möglichen Kalibrierungsmethoden und setzt zwei dieser Verfahren um. Dabei wird die Punkt-Kalibrierung (engl. Cross-wire Calibration) und die bildbasierte Kalibrierung aus der Publikation "Image-Based Method for in-Vivo Freehand Ultrasound Calibration" von Wolfgang Wein und Ali Khamene implementiert und ausgewertet.

Time: 09.04.2015, ca. 16:00
Location: Room 073 in the Fraunhofer IGD, Fraunhoferstrasse 5, S3|05
Speaker: Felix Gorschlüter (Advisor: Martin Hess)
Title: "Seamless Rendering für Augmented Reality auf Mobilgeräten" (Master thesis)
Abstract: Ein häufiges Problem von Augmented-Reality-Anwendungen ist, dass sich computergenerierte Overlays von der unterliegenden realen Szene abheben und so für den Benutzer kein konsistenter Gesamteindruck entsteht. Um diesem Problem entgegenzuwirken, soll ein Algorithmus entwickelt werden, welcher synthetische Bilder durch adaptives Color Balancing und Simulation von Motion Blur in Echtzeit so anpasst, dass sie sich plausibel in reale Szenen einfügen.

Time: 09.04.2015, ca. 15:45
Location: Room 073 in the Fraunhofer IGD, Fraunhoferstrasse 5, S3|05
Speaker: Yannick Drost, Florian Herrmann und Pascal Schardt (Advisor: Martin Hess)
Title: "Regionenbasiertes Feature Matching für repetitive Szenen" (Projektpraktikum - Capturing Reality)
Abstract: Eine Reihe an Fotografien des Frankfurter AfE-Turms kurz vor seiner Sprengung am 02.02.2014 soll für eine fehlerfreie grafische Rekonstruktion analysiert und aufbereitet werden. Die Herausforderung dieses Projektes zeigt sich anhand der gleichmäßigen Struktur und der sich wiederholenden Textur des zu untersuchenden Objektes. Für eine Zuordnung einzelner Bildpaare werden, in beiden Bildern vorkom- mende gleiche Merkmale gesucht. Bereiche des einen Fotos werden jedoch, auf Grund der ähnlichkeit nichtzugehörigen Bildbereichen des anderen Fotos zugeteilt, weshalb eine Registrierung des Datensatzes fehlschlägt und letztlich kein verwendbares Ergebnis entsteht. Aufgabe ist eine qualitativ bessere Rekonstruktion des Frankfurter AfE-Turms. Um die Problematik mit der unzulänglichen Rekonstruktionsqualität zu beheben, wird ein Ansatz zur Regionierung des Datensatzes implementiert um eine Voraussetzung für eine erfolgreiche Nachbildung zu liefern. Zu diesem Zweck werden die Bilder durch Abgrenzungen inhaltlich unterteilt, sodass den anschließend extrahierten Bildmerkmalen zusätzliche Informationen zur relativen Lage zur Verfügung stehen. Damit werden nur korrespondierende Bereiche innerhalb der Bilder für den Registrierungsansatz in Betracht gezogen.

Time: 09.04.2015, 15:00
Location: Room 073 in the Fraunhofer IGD, Fraunhoferstrasse 5, S3|05
Speaker: Andreas Muttscheller (Advisor: Martin Hess)
Title: "Augmented Reality in der Lagerlogistik" (Bachelor thesis)
Abstract: Die Bachelorarbeit wird das Thema Augmented Reality (AR) in der Lagerlogistik behandeln, genauer die Wegfindung eines Lagermitarbeiters zu einem Punkt im Lager. Zunächst wird die Positionsermittlung innerhalb des Lagers betrachtet. Wie kann man möglichst genau ein Gerät orten und welche Technologien sind dafür notwendig. Dafür werden unter anderem WLAN, Beacons, QR/AR Codes auf ihre Tauglichkeit untersucht. Auch der Blickwinkel sowie die Laufrichtungserfassung muss untersucht werden. Verschiedene Techniken, wie z.B. der Kompass in einem Gerät, werden daraufhin analysiert. Des Weiteren wird geprüft, wie sich die räumlichen Gegebenheiten in einem Lager auf die Funktechnologien (WLAN, Bluetooth) auswirken. Da die Lager und Regale größtenteils aus Metall gefertigt sind, muss deren Beeinflussung auf das Signal untersucht werden. Es werden außerdem verschiedene mobile AR SDKs untersucht, mit Hilfe deren die Informationen dem Nutzer angezeigt werden. Um das Lagerlayout und den Laufweg zu ermitteln wird die Navigation mit dem Lagerführungssystem LFS verbunden. Dort lässt sich bereits das Lagerlayout pflegen und auch die Wegberechnung anhand verschiedener Laufwegstrategien ausführen. Dieses Modul nennt sich "Transport Leit System" TLS und wird für die Wegberechnung genutzt. Für einen Prototyp der Navigationslösung wird der folgende Anwendungsfall betrachtet: Ein Lagermitarbeiter steht an einem beliebigen Punkt im Lager und soll an einen anderen Punkt geführt werden. Dieser Punkt wird vom LFS vorgegeben und ist in diesem Fall die nächste Pickposition bei der Kommissionierung. Die Anwendung ermittelt dabei die Position des Mitarbeiters und lässt den Laufweg über das TLS berechnen. Dieser wird dem Nutzer anschließend auf seinem Gerät angezeigt und aktualisiert, sobald er sich bewegt. Zum Ende wird zusammengefasst, ob die angestrebte Lösung produktiv einsetzbar ist.

Time: 09.04.2015, 14:00 - 14:45
Ort: Room 242 in the Fraunhofer IGD, Fraunhoferstrasse 5, S3|05
Speaker: Lukas Hermanns (Advisor: Tobias Franke)
Titlel: "Screen Space Cone Tracing for Glossy Reflections" (Bachelor thesis)
Abstract: Indirect lighting (also Global Illumination (GI)) is an important part of photo-realistic imagery and has become a widely used method in real-time graphics applications, such as Computer Aided Design (CAD), Augmented Realtiy (AR) and video games. Path tracing can already achieve photorealism by shooting thousands or millions of rays into a 3D scene for every pixel, which results in computational overhead exceeding real-time budgets. However, with modern programmable shader pipelines, a fusion of ray-casting algorithms and rasterization is possible, i.e. methods, which are similar to testing rays against geometry, can be performed on the GPU within a fragment (or rather pixel-) shader. Nevertheless, many implementations for real-time GI still trace perfect specular reflections only.
In this Bachelor thesis the advantages and disadvantages of different reflection methods are exposed and a combination of some of these is presented, which circumvents artifacts in the rendering and provides a stable, temporally coherent image enhancement. The benefits and > failings of this new method are clearly separated as well. Moreover the developed algorithm can be implemented as pure post-process, which can easily be integrated into an existing rendering pipeline. The core idea of this thesis has been presented as a poster at SIGGRAPH 2014 [Hermanns and Franke, 2014].

Time: 27.03.2015, 15:00
Location: Room 074 in the Fraunhofer IGD, Fraunhoferstrasse 5, S3|05
Speaker: Samir Aroudj (Advisor: Martin Hess)
Title: "Meshless Collision Detection for deformable Bodies" (Masterarbeit)
Abstract: Meshes are often used for simulation, rendering and collision detection. Deformable bodies allow strong deformations which can unfortunately lead to degenerated mesh elements. Melting, rip and fracture simulations are even more problematic due to potential topological body changes. Elaborate mesh recalculation called remeshing is necessary to realize topological changes and avoid such degenerated elements. Remeshing implementations are complex and thus prone to errors. Three different meshes for simulation, collision handling and rendering respectively must be adapted consistently in many applications causing a high computational load. Meshbased collision detection with its distinction between different elements such as vertices, edges and faces requires complex case differentiations which can induce unrealistic or unwanted behavior. A uniform concept to describe an intersection volume of two bodies is thus preferable. This work presents a volumebased approach for collision detection and handling. Together with already known, meshless simulation and rendering techniques, it forms a completely meshfree overall system. For collision detection, an additional point cloud is moved for each body by means of its simulated points. The intersection of the axis aligned bounding boxes of the two point clouds of a body collision pair is then determined. Afterwards, a voxel grid is created for this intersection which in turn is filled with the two point clouds. The volume gradient and penetrations depths are then cellwisely computed for this discretized, gridbased intersection volume. Voxel¬based, collision resolving penalty forces can be determined by means of these information. Lastly, these voxel forces are distributed to the simulated points of both deformable bodies. The resulting system is free of meshes, degenerated elements and remeshing. The pointbased concept facilitates topologically complex and fineresolution collision geometry. Furthermore, it is a simple and flexible basis for melting, rip and fracture simulations.

Time: 20.03.2015, 15:00
Location: Room 220 in the Fraunhofer IGD, Fraunhoferstrasse 5, S3|05
Speaker: Sebastian Windisch (Advisor: Thomas Gierlinger)
Title: "Interaktive punktbasierte Visualisierung großer oberflächenbezogener Simulationsdaten" (Master thesis)
Abstract: The distributed and computer-based simulation of physical phenomena generates data sets of sizes depending on the precision of the simulation, the considered time interval, and the size of the simulation domain. Large result data sets are difficult to visualize and analyze due to their data volume. This especially applies to interactive and explorative analysis employing conventional workstations which are widely used in the industry and research. In this scenario a user navigates through a visualization of a data set to locate relevant results. This is an application specific procedure to find e.g. unexpected data or to (dis-)prove a hypothesis. Often the result data is too large to be sent to a PC or to be processed locally. Limiting factors in this use case are the local main memory, the videocard memory, the computing power as well as the available network bandwidth for data transmission. There are some approaches which enable remote visualization of triangle-based models or scenes. However they either require long preprocessing times for the whole data set or the utilization of remote and local resources is strongly coupled, i.e. local navigation or modification of visualization parameters is not possible without using the remote resources. In this thesis an approach to the visualization of such data sets based on the works of Ge et al. [9] and Preiner et al. [17] is presented, prototypically implemented and evaluated. It performs a view dependent point-based sampling of a remote surface-located simulation data set. The incrementally obtained point cloud of simulation data is organized and temporarily stored in an octree which is synchronized between client and server. On the client the surface of the simulation data set is then reconstructed by a splat-based reconstruction method. With the aid of deferred rendering and a color mapping local modifications of the simulation data visualization are possible. This approach allows a decoupling of the data complexity and rendering performance between client and server. The viability of the method is experimentally evaluated using a small sample data set. An evaluation for large data could not be conducted due to the lack of available data sets. Based on the work of Ge et al. [9] it can be assumed that there is a direct dependency between the amount of data that can be visualized at interactive framerates and the available server-based resources.

GRIS Kolloquium

Time: 26.02.2015, 15:00
Location: Room 074 in the Fraunhofer IGD, Fraunhoferstrasse 5, S3|05
Speaker: Prof. Dr.-Ing. Soergel
FG Fernerkundung und Bildanalyse, TU Darmstadt
Title: "Trends in Remote Sensing: Modern Sensors and Analysis Methods"
Abstract: General trends in remote sensing are the improvements of geometrical and spectral resolution of sensors as well as time-series analysis and data fusion. After a general discussion of such trends the talk will focus in more detail on the active techniques Synthetic Aperture Radar and Airborne Laserscanning. In the former we witness an enormous success of techniques which rely on interferometric processing, for example, the ongoing TanDEM mission and subsidence monitoring by time series analysis. In the field of laserscanning there are some interesting trends too, for instance, bathymetry and so-called multi-spectral laser.
With respect to analysis methods two methods are discussed namely Compressive Sensing and Conditional Random Fields
Bio: *since 10/2013 Full Professor for Remote Sensing and Image Analysis at Technische Universität Darmstadt
*01/2010-09/2013 Associate Professor for Radar Remote Sensing and active Systems at LUH.
*01/2006-12/2009 Assistant Professor for Radar Remote Sensing at the Institute of Photogrammetry and GeoInformation (IPI) of LUH, Germany.
*03/2003 Doctorate at the department of Electrical Engineering and Information Technology of Leibniz Universität Hannover (LUH), Germany.
*05/1997-12/2005 Research associate at the Institute for Optronics and Pattern Recognition (FOM) in Ettlingen, Germany.
*11/1990-03/1997 Diploma in Electrical Engineering, University of Erlangen, Germany.

Thesis presentations

Time: 25.02.2015, 15:00
Location: Room 072 in the Fraunhofer IGD, Fraunhoferstrasse 5, S3|05
Speaker: Robert Cibulla (Advisor: Tatiana von Landesberger)
Title: "Visuelle Analyse von Auswirkungen von Verspätungen im Zugverkehr" (Bachelor thesis)

Time: 11.02.2015, ca. 13:40
Location: Room 324 in the Fraunhofer IGD, Fraunhoferstrasse 5, S3|05
Speaker: Johannes Merz (Advisor: Tatiana von Landesberger)
Title: "Analyse und Visualisierung von Mesh-Korrespondenzen hinsichtlich der Foldover-Problematik" (Bachelor thesis)

Time: 11.02.2015, 13:00
Location: Room 324 in the Fraunhofer IGD, Fraunhoferstrasse 5, S3|05
Speaker: Chridtoph Niese (Advisor: Tatiana von Landesberger)
Title: "Visualisation of composer relationships using implicit data graphs" (Bachelor thesis)

Guest lecture in PMPP - Programming Massively Parallel Processors

Time: 10.02.2015, 11:30 - 13:20
Location: Room 074 in the Fraunhofer IGD, Fraunhoferstrasse 5, S3|05
Speaker: Dr. Michael Klemm
Senior Application Engineer
Software and Services Group, Developer Relations Division,
Intel GmbH
Title: "Programming for the Intel Xeon Phi Coprocessor"
Abstract: The Intel(R) Xeon Phi(tm) Coprocessor Architecture and Programming. The Intel Xeon Phi Coprocessor is a massively parallel compute device to provide additional compute power for traditional compute servers. It is designed to resemble the flexibility and programmability of Intel Architecture, while providing a large number of threads and extra-wide SIMD instructions. In this presentation, we present the coprocessor architecture from a hardware and software perspective. We discuss the tools and programming models available. While some of the well-known parallel programming models are touched, the focus is on the coprocessor-specific extensions of these programming models. We present a case study that shows how the concepts shown earlier are tied together in a quantum chemistry application scaling to more than 60,000 cores.
Bio: Dr. Michael Klemm is part of Intel's Software and Services Group, Developer Relations Division. His focus is on High Performance and Throughput Computing. Michael received a Doctor of Engineering degree (Dr.-Ing.) in Computer Science from the Friedrich-Alexander-University Erlangen-Nuremberg, Germany. His research focus was on compilers and runtime optimizations for distributed systems. Michael's areas of interest include compiler construction, design of programming languages, parallel programming, and performance analysis and tuning. Michael is Intel representative in the OpenMP Language Committee and leads the efforts to develop error handling features for OpenMP.
If you are interested in meeting with Dr. Michael Klemm after the lecture, please send an email to:

GRIS Kolloquium

Time: 23.01.2015, 15:30
Location: Room 074 in the Fraunhofer IGD, Fraunhoferstrasse 5, S3|05
Speaker: Rafael Spring
DotProduct LLC
Title: "Mobile 3D Sensing at DotProduct - Applications and Future"
Abstract: 3D cameras are making their way into phones and tablets. Similar to traditional mobile sensors such as 2D cameras, GPS and IMUs miniaturization and mass production have allowed them to go from external accessory to integrated and cheap. At the same time mobile computational power has exploded and is now approaching console-level performance in the current generation of SoCs. In this talk I will describe what all this means for computer vision, how it will affect consumer and professional markets on a high-level, what will be the killer-applications and how DotProduct technology helps enable them. I will also briefly talk about some upcoming trends in 3D sensing and will show some live demos of our products.
Bio: Rafael Spring is an ex-Googler and computer vision engineer. He was part of the Google Visual Search Team in Santa Monica from 2009 to 2011, after his project "Enkin" (an Augmented Reality navigation system for Android) was acquired by Google in 2008. During his time at Google he developed real-time computer vision algorithms (2D) and contributed to Google Goggles,Google Glass and the early project Tango.
He is now founder and CTO of DotProduct and wants to make realtime 3D perception an everyday capability of mobile devices.

Thesis presentations

Time: 23.01.2015, 14:15-15:00
Location: Room 074 in the Fraunhofer IGD, Fraunhoferstrasse 5, S3|05
Speaker: Daniel Thuerck (Advisors: Michael Goesele, Marc Pfetsch, Sven Widmer)
Title: "Optimizing large-scale irregular Markov Random Fields on GPUs" (Master thesis)
Abstract: Energy optimization by graph cuts in $\alpha$-Expansion have become ubiquitous in computer vision during the last decade. Especially for grid topologies, fast multicore and even massively-parallel solvers have been developed. However, those are based on cache-optimal storage of the topology and hardcoded neighborhoods. For irregular topologies, few implementations of different methods are available. Additionally, none of them uses multicore systems or even massively-parallel systems such as graphics processing units (GPUs). In this thesis, we first review the state-of-the-art techniques and analyze their potential for parallelization. After pointing out that most methods are inherently serial, we present two novel methods for energy minimization in pairwise markov random fields: (Heuristic) iterative monotonic fixing (HIMF) and monte carlo partial cuts (MCCO). While HIMF adapts the method of dynamic programming with different heuristics, MCCO relies on a theoretical foundation to find a minimum s-t-cut in a graph without computing the maximum flow. Theorems that state sufficient conditions for partially solving those cuts are given and for the first time proven. Both methods are evaluated against different (single core) author implementations that are available.

Time: 23.01.2015, 13:30-14:15
Location: Room 074 in the Fraunhofer IGD, Fraunhoferstrasse 5, S3|05
Speaker: Stepan Konrad (Advisors: Michael Goesele, Simon Fuhrmann)
Title: "Pose Estimation and Loop Closing from Video Data" (Bachelor thesis)
Abstract: In robotics the simultaneous localisation and mapping (SLAM) algorithms are a well studied approach to estimate the position of a robot vehicle while creating a map of the surrounding. The majority of these algorithms use odometry or GPS sensors to cope with large outdoor trajectories. From a similar point of view the computer vision community uses structure from motion (SfM) algorithms to estimate accurate camera poses of an unconstrained image data set. In the past few years the video resolution of consumer cameras has reached a level where it becomes attractive for research purposes as input to these algorithms. The goal of this thesis is to adapt an SfM approach to use this video data. However there are two main problems: The approach has to handle a large number of input frames efficiently while still detecting similar previously seen locations (loops) of the input data without performing an exhaustive matching of all image pairs. This thesis presents an approach using a vocabulary tree guided matching scheme which solves this problem. Performance is compared to exhaustive matching on different input scenes.
However, this is still not sufficient to reconstruct large datasets that contain loop closures in the camera path. Due to the incremental manner of the majority of SfM algorithms, drifts occur during the estimation of camera poses. In this thesis different solutions to this problems are discussed. One specific solution using a global bundle adjustment with additional loop closing constraints is demonstrated on a large outdoor scene containing multiple loops.

GRIS Kolloquium

Time: 21.01.2015, 11:00
Location: Room 074 in the Fraunhofer IGD, Fraunhoferstrasse 5, S3|05
Speaker: Daniel Gritzner
Title: "3D Graphics Rendering using a Polygon-Voxel-Hybrid Approach"
Abstract: There are various approaches to rendering 3D graphics on a computer. The two most common approaches are rasterization and ray tracing. Each approach has different advantages and disadvantages, e.g. rasterization is generally fast while ray tracing is able to deliver a high level of realism. There are also different kinds of primitives to describe 3D scenes, the most common ones being polygons and voxels. Polygons describe surfaces while voxels describe volumes. Again, there is no single type of primitive which is optimal for every application.
The goal of this presentation is to show how rasterization of polygons and ray tracing of voxels can be combined into a single renderer which is able to render each object in an optimal way based on its properties. The reason for doing so is to work towards a renderer which gains as many advantages of each approach as possible. The presentation discusses problems one faces when combining the previously mentioned techniques, such as showing each object at the correct depth, and the viability of this idea by examining benchmarks of an actual implementation.

Time: 28.11.2014, 10:00
Location: Room 074 in the Fraunhofer IGD, Fraunhoferstrasse 5, S3|05
Speaker: Prof. Douglas Cunningham
Department for Graphical Systems, TU Cottbus
Title: "A brief introduction to Perceptual Experiments"
Abstract: For over 150 years, physiologists and psychologists have been performing experiments to determine what signals in the world humans and animals can extract, how those signals are converted into information, and how the information is then represented and processed. Recently, there has been an increasing trend of computer scientists performing similar experiments, although often with quite different goals. While at first glance such experiments seem to be easy to design, carry out, and analyze, a century and a half of experience has shown that there are numerous hidden traps which can waylay the unaware. This talk will briefly introduce all the basic elements of a single experiment and will derive a mathematical description of experimental design. In the process it will become clear why precision and balance must lie at the core of any experiment.
Bio: Douglas W. Cunningham received his Ph.D in psychology from Temple University in 1997. He is currently leads the graphical systems department at the Brandenburg Technical University. Before becoming a professor at BTU, he was a postdoctoral researcher at the University of Tübingen, at the MPI for Biological Cybernetics, and for Logicon Technical Services. His research interests focus on integrating psychology and informatics, and include topics such as facial expressions, image statistics, sensorimotor adaptation, and computational aesthetics. In 2013 he published a book together with Erik Reinhard and Tania Pouli called "Image Statistics in visual computing", in which they examine the fundamentals of perceptually relevant image statistics and how they have been applied in computer graphics. The basis for this talk is the first few chapters of a book he published in 2011 called "Experimental design: from user studies to psychophysics" together with Christian Wallraven.

Thesis presentations

Time: 24.11.2014, 10:00-10:30
Location: Room 073 in the Fraunhofer IGD, Fraunhoferstrasse 5, S3|05
Speaker: Sebastian Zander-Walz (Betreuer: Andreas Braun)
Title: "Curved large-area surfaces for gestural interaction" (Bachelorarbeit)
Abstract: Gestures are a natural and intuitive part of human communication. Since the appearance of smartphones and tablet computers, gestural interaction became suitable to many customers. Usually gesture interaction is implemented using two dimensional planar surfaces, although the natural movement of the human body results in elliptic or spherical paths. This thesis shows a way of equipping large-area curved surfaces with capacitive loading-mode proximity sensors and gesture recognition from these sensors data. Therefore already existing techniques, well-known from the use in planar system, were adapted to the use in curved prototypes. To prove the results both, the interaction with the prototype and the gesture recognition have been evaluated and the results discussed.

Time: 21.11.2014, 14:00-14:30
Location: Room 073 in the Fraunhofer IGD, Fraunhoferstrasse 5, S3|05
Speaker: Stefan Krepp (Advisor: Andreas Braun)
Title: "Unsichtbare Erkennung von Berührungen über akustisches Tracking" (Master thesis)
Abstract: Die unsichtbare Integration von Eingabegeräten in die natürliche Umgebung der Nutzer ist eine große Herausforderung, die allerdings große Chancen bietet. Allgegenwärtige Eingabegeräte besitzen das Potential gewohnte Interaktionsmuster zu durchbrechen. Eingabegeräte wie klassische Fernbedienungen für Fernseher könnten bald der Vergangenheit angehören. Ein Türöffner für unsichtbare Benutzerschnittstellen sind akustische Sensoren, die sich kostengünstig und einfach installieren lassen und fast jede Oberfläche zu einem Eingabegerät machen können. In dieser Arbeit wurde ein akustisches Berührungserkennungssystem entwickelt, das die Erkennung und Unterscheidung verschiedener Impact-Gesten und Wischgesten ermöglicht und sich unsichtbar in bestehende Möbel integriert. Dabei wurde ein bestehender Ansatz erweitert. Erste Nutzerstudien unterstreichen die Machbarkeit des verfolgten Ansatzes. Zudem konnte gezeigt werden, dass akustische Berührungserkennung einen Mehrwert für andere Technologien bedeuten kann. So konnte gezeigt werden, dass die Interaktionsgeschwindigkeit eines kapazitiven Gestenerkennungssystems durch die Integration akustischer Berührungserkennung erhöht werden kann.

Time: 18.11.2014, 16:00-16:30
Location: Room 073 in the Fraunhofer IGD, Fraunhoferstrasse 5, S3|05
Speaker: Patrick Lerch (Advisor: Tatiana von Landesberger)
Title: "Layoutalgorithmus fuer Visualisierung von Mustern in Ausfalleffektgraphen" (Bachelor thesis)

GRIS Kolloquium

Time: 18.11.2014, 10:00
Location: Room 073 in the Fraunhofer IGD, Fraunhoferstrasse 5, S3|05
Speaker: Dr. Daniel Kondermann
Heidelberg Collaboratory for Image Processing
Title: "Ground Truth for Correspondence Problems"
Abstract: Creating ground truth for e.g. optical flow in natural outdoor environments seems almost impossible. In this talk, I will propose several possible approaches, some of which we are currently investigating. One approach is to use semi-automatic vision algorithms as is done for example in movie postproduction to create 'weak' ground truth. Another method is to evaluate the properties of today's computer graphic rendering systems with respect to their ability to generate images close to the real world. Crowdsourcing is yet another interesting approach.
I will also discuss the problems of content selection, defining performance measures and benchmarking with respect to correspondence estimation and related algorithms.

Thesis presentations

Time: 13.10.2014, 13:00-13:30
Location: Room 072 in the Fraunhofer IGD, Fraunhoferstrasse 5, S3|05
Speaker: Olav Lenz (Advisor: Tatiana von Landesberger)
Title: "Visueller Vergleich von Ausfall-Effekten in mehreren Datensätzen" (Master thesis)

Time: 26.08.2014, 09:00-09:30
Location: Room 072 in the Fraunhofer IGD, Fraunhoferstrasse 5, S3|05
Speaker: Julia Hilpert (Advisor: Florian Jung)
Title: "Halbautomatische Segmentierung von Lymphknoten aus Magnetresonanztomographiedaten des Kopf- und Halsbereichs" (Bachelor thesis)
Abstract: Knowledge of size and appearance of lymph nodes in cancer patients can allow a better understanding of the spread of a tumour. Therefore, measuring the lymph nodes is important for cancer treatment as well as cancer prevention. As a first step a segmentation of lymph nodes is necessary.
So far there are few approaches on automated MRI lymph node segmentation. This thesis presents a method used to segment lymph nodes from MRI scans of the head and neck area. The segmentation requires minimal user interaction by simply pickinga point within the lymph node tissue. The main segmentation method used is a watershed transformation, which segments the image according to its gradient. Radial ray based segmentation allows an approximation of the surface area of the lymph node.
This surface serves to determine, among the images obtained, the segments representing the lymph node. The method was evaluated on 95 lymph nodes from 17 different T1-weighted MRI data sets. The Dice Similarity Coefficient was at 0.69±0.15. As a result, the developed method lays the foundation to a fully automated segmentation of lymph nodes.

Time: 19.08.2014, 16:30-17:00
Location: Room 220 in the Fraunhofer IGD, Fraunhoferstrasse 5, S3|05
Speaker: Kevin Daun (Advisor: Daniel Weber)
Title: "Collision Handling between Rigid and Deformable Bodies withContinuous Penalty Forces" (Bachelor thesis)

Time: 18.08.2014, 15:00-15:30
Location: Room 073 in the Fraunhofer IGD, Fraunhoferstrasse 5, S3|05
Speaker: Manuel Kopp (Advisor: Tatiana von Landesberger)
Title: "Methods and visual-interactive tools for automatic generation of geo-located graphs. Application to supply chain networks." (Master thesis)

GRIS Kolloquium

Time: 28.07.2014, 16:00
Location: Room 074 in the Fraunhofer IGD, Fraunhoferstrasse 5, S3|05
Speaker: Dr.-Ing. Jörn Kohlhammer
Fraunhofer IGD
Title: "Die Visual Analytics-Forschung in Zeiten von Big Data"
Abstract: Visual Analytics ist die Verbindung von Informationsvisualisierung und automatischen Methoden mit dem Ziel einer besseren Entscheidungsfindung.
In den letzten 10 Jahren hat sich dieses Gebiet als ein wesentlicher Trend in der Visualisierungsforschung etabliert. Das optimale Zusammenspiel automatischer Ansätze mit Visualisierungsmethoden stellt dabei eine Herausforderung dar, der man sich in jedem Anwendungsgebiet besonders stellen muss. Denn gerade bei massiven Datenmengen, wie sie aktuell unter dem Schlagwort Big Data subsumiert werden, ist neben den zu visualisierenden Daten eine Betrachtung der Aufgabe und des Nutzers von großer Bedeutung. Der Vortrag geht auf aktuelle Ansätze in der Visual Analytics-Forschung ein und zeigt anhand von Beispielen aus verschiedenen Anwendungsgebieten, wie sich die Vorteile von Visual Analytics realisieren lassen. Der Vortrag endet mit einem Überblick über die neuesten Entwicklungen im Bereich Visual Analytics, auch im Hinblick auf eigene Forschungsziele.

Time: 22.07.2014, 10:00
Location: Room 074 in the Fraunhofer IGD, Fraunhoferstrasse 5, S3|05
Speaker: Dr. Oliver Grau
Intel Visual Computing Institute
Title: "Interactive 3D content = new user experiences?"
Abstract: This presentation reviews some of the technological aspects in visual computing that look into scanning, processing and displaying of 3D content. Many of the techniques are currently fueled by advances in mobile computing and allow capture of highly detailed geometry and appearance of real objects. Further, progress is made on client-server technologies and display techniques, with a revival of AR and VR. The question that is not yet sufficiently answered is: What is in for the user?

Bio: Oliver Grau jointed Intel in 2012 as one of the directors of the Intel-Visual Computing Institute in Saarbrücken. Before that he worked more than a decade for BBC R&D in London on Computer Vision projects for innovative production techniques for studio, sport and interactive.

Time: 10.07.2014, 10:00
Location: Room 074 in the Fraunhofer IGD, Fraunhoferstrasse 5, S3|05
Speaker: Prof. Pohl
Human Computer Iteraction Group, TU Wien
Title: "Sensemaking with Visual Analytics"
Abstract: Visual analytics is the science of analytical reasoning facilitated by interactive visualization. This implies that human reasoning processes play a major role. The design of visual analytics systems should support these reasoning processes. Unluckily, we do not know enough about these processes and how they are related to the design of visualizations.
Approaches from cognitive psychology can be used to clarify this issue.
One possibility to explain reasoning processes based on visual analytics systems is the sensemaking theory as formulated by Gary Klein. Klein emphasizes the importance of restructuring one's mental models to achieve insights. To test which theoretical approach is able to explain reasoning with visual analytics systems the interaction processes with such systems have to be analyzed. The ultimate goal is to find design guidelines for an appropriate design of visual analytics systems.

Thesis presentations

Time: 10.07.2014, 11:00-11:30
Location: Room 074 in the Fraunhofer IGD, Fraunhoferstrasse 5, S3|05
Speaker: Martin Distler (Advisor: Tatiana von Landesberger)
Title: "Ähnlichkeitswahrnehmung von Graphen mit beschrifteten Knoten in einer Node-Link Darstellung" (Master thesis)
Abstract: Die vorliegende Arbeit untersucht die Wahrnehmung von Ähnlichkeit zwischen Node-Link-Diagrammen mit beschrifteten Knoten. Ziel ist es hier herauszufinden, worauf Betrachter bei dem Vergleich zwischen Node-Link-Diagrammen achten und ob sich die betrachteten Merkmale zwischen verschiedenen Betrachtern unterscheiden. Anhand dieser Erkenntnisse wäre es somit möglich eine Ähnlichkeitsmaß zu erstellen, das sowohl die strukturellen als auch inhaltlichen Eigenschaften der Graphen berücksichtigt. Hierzu wurde eine Studie entwickelt bei der 24 Probanden gebeten wurden die Ähnlichkeit von insgesamt 48 Graphen-Paaren zu bewerten und die betrachteten Merkmale zu nennen. Die Probanden teilten sich dabei in zwei Expertengruppen (Informatiker & Psychologen) auf. Zu den Erkenntnissen zählt der besondere Einfluss eines Kontextknotens und seine Position im Verhältnis zu den restlichen Knoten des Graphen. Zudem gab es einen Überlagerungseffekt durch optisch dominantere Veränderungen und führte demnach zu einer Bewertung der Ähnlichkeit aufgrund falscher Betrachtungen.

Time: 11.07.2014, 15:00-15:30
Location: Room 073 in the Fraunhofer IGD, Fraunhoferstrasse 5, S3|05
Speaker: Yannick Drost (Advisor: Wissam El Hakimi)
Title: "Bildrekonstruktion aus Kegelstrahlprojektionen für zwei Scan-Trajektorien" (Bachelor thesis)
Abstract: C-Bogen Geräte spielen im humanmedizinischem Bereich derzeit sowohl für verbessert diagnostische Planungen als auch bei intraoperativen Anwendungen eine immer wichtigere Rolle. Wegen ihrer hochauflösenden Bildgebung wird eine CBCT-Anlage für die Bildrekonstruktion eines Untersuchungsobjektes genutzt. Die aktuelle Aufnahmetrajektorie ist für das verwendete kegelstrahlbasierte C-Bogensystem eine einzelne Bogentrajektorie. Zudem ist der Rotationsbereich des C-Bogen Gerätes auf unter 180◦ beschränkt, sodass akquirierte Projektionsbilder nicht für eine akzeptable Bildqualität des zu rekonstruierenden Bildes ausreichen.
Im Rahmen dieser Arbeit wurde für mit einem Flächendetektor ausgestatteten mobilen C-Bogen CT-Gerät (Ziehm Vision RFD) eine Kombination von zwei separaten Projektionstrajektorien entwickelt, die ausreichend Rohdaten für eine aussagekräftigere Bildrekonstruktion liefern soll. Wegen des störenden Artefaktverhaltens und des durch fehlende Projektionen reduzierten Informationsgehalts werden Markierungen am Untersuchungsobjekt angebracht, um den Prozess einer präzisen Registrierung der einzelnen Projektionsdatensätze zu ermöglichen. Die verwendete Rekonstruktionsmethode ist eine Erweiterung eines iterativen Verfahrens für zwei Kreisbogenbahnen, welche die Daten entsprechend gewichtet und das Artefaktverhalten reduziert.

Time: 11.07.2014, 16:00-16:30
Location: Room 073 in the Fraunhofer IGD, Fraunhoferstrasse 5, S3|05
Speaker: Max Mehltretter (Advisor: Wissam El Hakimi)
Title: "Optimierung einer 2D/3D Registrierungsmethodik für bildgestützte Eingriffe" (Bachelor thesis)
Abstract: In der modernen Medizin spielen bildgebende Verfahren eine immer größere Rolle. Sie helfen minimalinvasive Eingriffe zu ermöglichen, wodurch es zu einer Reduktion des daraus resultierenden Traumas für den Patienten kommt. Ein bereits häufig verwendetes Verfahren ist die 2D-3D-Registrierung. Es registriert ein präoperativ aufgenommenes 3D-Volumen mit einem intraoperativ gewonnen 2D-Bild. Das Resultat ist eine möglichst exakte Überlagerung beider Bilddaten und kann für verschiedene Anwendungszwecke verwendet werden. Neben der Erstellung eines Differenzbildes erlaubt es dem Operateur auch die Orientierung im drei dreidimensionalen Raum, obwohl ihm nur eine aktuelle 2D-Aufnahme vorliegt. Die hier vorgestellte Arbeit beschäftigt sich mit einer solchen Registrierungsmethodik, die speziell auf Bilddaten des Schädels Anwendung findet. Ein großer Teil, der entwickelten Verfahren bietet bereits die erforderliche Genauigkeit, benötigt dafür aber selbst auf leistungsstarker Hardware oftmals mehrere Minuten für die Registrierung eines einzelnen 2D-Bildes. Da während einer Operation jedoch immer aktuelle Bilddaten zur Verfügung stehen müssen, ist eine solche Laufzeit inakzeptabel und macht das Verfahren für die Praxis untauglich. Das Ziel dieser Arbeit ist es daher, das zu Grunde liegende Verfahren zu optimieren und die Laufzeit damit zu verkürzen. Zu diesem Zweck werden verschiedene Optimierungsansätze vorgestellt und analysiert. Neben der Untersuchung eines kombinierten Multi-Resolution- und Multi-Skalen-Ansatzes, liegt das Hauptaugenmerk auf der Portierung zeitintensiver Komponenten auf die GPU. Realisiert wird dies durch Anwendung der GPGPU-Lösung CUDA. Es wurde gezeigt, dass damit eine massive Reduktion der Laufzeit möglich ist, ohne dass sich die Genauigkeit der Registrierung verschlechtert.

Time: 14.07.2014, 09:30-10:00
Location: Room 072 in the Fraunhofer IGD, Fraunhoferstrasse 5, S3|05
Speaker: Alexander Geurts (Advisor: Tatiana von Landesberger)
Title: "Visueller Vergleich von Segmentierungsalgorithmen von medizinischen 3D-Bildern" (Bachelor thesis)
Abstract: Diese Bachelorarbeit bietet einen neuen Ansatz für das Vergleichen mehrerer automatisch segmentierender Algorithmen von medizinischen 3D-Bildern. Dabei werden von Experten manuell segmentierte Referenz-Daten als Ground Truth verwendet. Nach Eingabe von zu testenden Samples werden diese global, lokal und regional evaluiert. Dabei werden mehrere Algorithmen miteinander verglichen und Empfehlungen ausgegeben, welcher Algorithmus im Allgemeinen oder in einzelnen Regionen zum Segmentieren benutzt werden sollte. Das hierbei erstellte Programm bietet dabei sowohl interaktiv-visuelle als auch analytische Sichten auf die zu evaluierenden Samples. Die Entwickler der Algorithmen können mit diesen Informationen ihre Algorithmen evaluieren und verbessern, was bei der automatischen Segmentierung von medizinischen Bildern sehr hilfreich sein kann. Das im Rahmen dieser Arbeit erstellte Programm ist dabei auf den Vergleich von bis zu sechs Algorithmen ausgerichtet, kann jedoch prinzipiell auch für den Vergleich von mehr Datensätze gleichzeitig benutzt werden.

Thesis presentations

Time: 17.06.2014, 11:45-12:15
Location: Room 073 in the Fraunhofer IGD, Fraunhoferstrasse 5, S3|05
Speaker: Beatrice Frieß (Advisor: Wissam El Hakimi)
Title: "Automatische Bestimmung der optimalen Parameter für intraoperative Bildaufnahmen und Instrumentenverifikation" (Master thesis)
Abstract: Die Multi-Port-Knochenchirugie ist ein minimal invasives Verfahren, dessen Ziele unter anderem eine minimal traumatische Operationsprozedur und geringere Morbidität sind. Bei diesem Verfahren werden drei Bohrungen in den Schädel durchgeführt, die durch mehrere CBCT Aufnahmen verifiziert werden. Diese Aufnahmen sind für den Patienten mit zusätzlicher Strahlungsbelastung verbunden und erfordern einen höheren Zeitaufwand für den Eingriff. Das Ziel dieser Arbeit ist die Unterstützung des Operators bei der Durchführung der intraoperativen Aufnahmen durch den Vorschlag geeigneter Aufnahmeparameter. Weiterhin erfolgt eine automatische Berechnung der Abweichung, der gebohrten Kanäle, von der Planung. Dafür wird die Bohrermittelinie auf zwei intraoperativen Projektionsbilder automatisch bestimmt, und durch Rückprojektion, die aktuelle 3D Lage des Bohrers bestimmt. Anschließend erfolgt eine grafische Visualisierung der Abweichung, um dem Anwender eine räumliche Darstellung zu bieten. Eine Evaluation dieser Arbeit zeigt, dass durch den automatischen Vorschlag optimaler Aufnahmeparameter die Akquisition unbrauchbarer Aufnahmen vermieden werden kann, und somit die unnötige Strahlungsbelastung.

GRIS Kolloquium

Time: 11.06.2014, 15:30
Location: Room 074 in the Fraunhofer IGD, Fraunhoferstrasse 5, S3|05
Speaker: Prof. Rothkopf
Institute of Psychology, Technische Universitaet Darmstadt
Title: "What guides vision during natural behavior?"
Abstract: The allocation of gaze has commonly been studied empirically under a variety of laboratory tasks that differ quite drastically from vision in naturalistic, extended, sequential tasks. Theoretical models of gaze allocation have been based largely on the saliency hypothesis, which posits that human gaze is directed towards regions of a scene that are high in contrast within a few specific feature dimensions. How well do these results transfer to vision in every-day naturalistic behavior?
We provide evidence from a series of studies and from computational modelling that low level image features such as saliency are weak predictors of gaze under naturalistic behavior and show that instead behavioral goals within a task and specific task contexts exert a far great influence on eye movement behavior. Surprisingly, inter-subject variability in such naturalistic tasks is quite small, when considering specific state variables that are relevant for representing behavioral goals. We quantify these goals as intrinsic preferences through Bayesian generative models of behavior and by explicitly manipulating the cost structure of tasks through a gaze contingent paradigm we demonstrate the adaptability of the visual system to these costs.

Thesis presentations

Time: 26.05.2014, 09:00-09:30
Location: Room 073 in the Fraunhofer IGD, Fraunhoferstrasse 5, S3|05
Speaker: Tobias Haberstich (Advisor: Prof. Dr.-Ing.Georgios Sakas)
Title: "Automatische Fusion von intraoralen Röntgenaufnahmen mit DVT Volumen auf Basis einer 2D/3D Registrierung" (Master thesis)
Abstract: Bildgebende Verfahren sind im dentalen Bereich von großem Nutzen. Von der Kariesdetektion, der dentalen Biometrie zur Identifikation, bis hin zur Planung von Wurzelbehandlungen sind viele weiterführende Anwendungen mithilfe von Algorithmen möglich. In dieser Arbeit wird eine Methode vorgestellt, um Datensätze aus der dentalen Bildgebung miteinander zu fusionieren. Diese sind hier speziell 2D Intraoralaufnahmen und 3D DVT Volumen. Dies entspricht einer 2D/3D Registrierung aus der medizinischen Bildverarbeitung. Es wird eine automatische Initialisierung vorgestellt, die ohne Kenntnis der realen Anordnung einer Intraoralaufnahme eine solche Registrierung ermöglicht. Des Weiteren werden Komponenten gezeigt, die typische Störfaktoren einer multimodalen Registrierung kompensieren. Zuletzt werden Möglichkeiten für eine gemeinsame Darstellung der Daten gegeben.

GRIS Kolloquium

Time: 25.04.2014, 09:30
Location: Room 074 in the Fraunhofer IGD, Fraunhoferstrasse 5, S3|05
Speaker: Dr. Stefan Guthe
Department of Computer Science, TU Clausthal
Titel: "Perceptual Graphics"
Abstract: Traditionally computer graphics is about finding the most suitable representation for any given scene or scenario. Especially in the area of image synthesis, this leads to trying to model physical processes as exact as possible. However, usually analytic solutions are not possible and numerical approximations require a huge number of computations to converge. Luckily, even though a physically correct image is also perceptually correct, a perceptually correct image doesn't have to be physically accurate. In perceptual graphics, one of the open questions is, what makes an image look real and how can this realism be quantized.
With this question in mind, we look at some problems that try to model real world objects in computer graphics. For this we use a variety of different sensors like regular cameras and active light sensors. Since most reconstruction problems are not well defined, we need to supply additional information about what a viewer expects the object to look like. As perceptual graphics is about images or videos, we are interested in how the images or videos of these reconstructed objects are perceived. In order to quantify the perception, one has to take the human visual system into account. Instead of taking the algorithmic approach trying to model these processes, we directly use the human visual system. For this, we capture EEG data and try to find correlations between perceived quality and activity in the brain. This approach is the so called human in the loop.
The vision behind all this is to define and quantize what makes an image real and what short cuts can be take while generating images without sacrificing both realism and perceived quality. Other interesting questions are if we can use the human in the loop to optimize specific parameters during image generation or if we can actually take the human out of the loop again without modelling the whole visual system.

Thesis presentations

Time: 01.04.2014, 13:00-13:45
Location: Room 073 in the Fraunhofer IGD, Fraunhoferstrasse 5, S3|05
Speaker: Katharina Eckeren (Advisor: Reimar Tausch, PD Dr. Arjan Kuijper, Prof. Dr.-Ing. Michael Goesele)
Title: "Entwicklung und Evaluation eines 3D Object Retrieval Systems für digitalisierte Kulturobjekte" (Master thesis)
Abstract: This thesis is about the construction, implementation and evaluation of a 3D object retrieval system for digitized cultural objects. After the integration and introduction to the used description method for 3D objects, called 3D Histogram of Oriented Gradients (abbr. 3D-HOG) [1][2] into the context of the current state of the art an explanation of each component of the retrieval system follows. In addition to the detailed description of the functional principle of the 3D-HOG descriptor a data basis was constructed. The data basis consists of cultural objects with high resolution polygon meshes. The evaluation contemplated different parameter choices and uses the measure of the R-precision [3] for early grading. By using the block normalization, the cell quantity and the size of the gradient field, the three most important parameters of 3D-HOG, it was possible to improve the R-precision from 0.799 to 0.888. Furthermore it was shown that the resolution of the polygon meshes is a matter of secondary importance as the resolution barely influences the results of the chosen description method.

Time: 01.04.2014, 13:45-14:30
Location: Room 073 in the Fraunhofer IGD, Fraunhoferstrasse 5, S3|05
Speaker: Nils Möhrle (Advisor: Prof. Dr.-Ing. Michael Goesele)
Title: "Texturing 3D Reconstructions from High-Resolution Multi-Scale Images" (Bachelor thesis)
Abstract: Texturing is typically the last step within the pipeline for 3D object reconstruction. In the last decade image based 3D reconstruction algorithms became robust and efficient enough to reconstruct even large and unconstrained scenes. However, texturing algo- rithms struggle with such datasets. This work introduces an algorithm that creates a high quality texture for the resulting meshes of such datasets. Based on Lempitsky and Ivanov's work [19] we use graph cuts to simultaneously select an appropriate view for each face and minimize visible seams. Within the view selection we account for sharp- ness, resolution and distance of the view as well as the viewing angle. In order to handle the remaining visible seams we use a global seam leveling procedure to adjust color val- ues of the texture patches. Apart from this adjustment we do not blend or resample the input images in any way such that the resulting texture patches have almost identical quality compared to the input images. The algorithm handles datasets with more than 500 high resolution images and meshes with over 8 million triangles in a few hours.

Time: 01.04.2014, 14:30-15:15
Location: Room 073 in the Fraunhofer IGD, Fraunhoferstrasse 5, S3|05
Speaker: Stepan Konrad (Advisor: Prof. Dr.-Ing. Michael Goesele)
Title: "Pose Estimation and Loop Closing from Video Data" (Current status Bachelor thesis)

Time: 28.03.2014, 10:00-10:30
Location: Room 048 in the Fraunhofer IGD, Fraunhoferstrasse 5, S3|05
Speaker: Elisa Garica Corisco (Advisor: Arjan Kuijper)
Title: "Registration of Ultrasound medical Images" (Master thesis)
Abstract: Ultrasound imaging is a widely used medical diagnostic technique due to its noninvasive and low cost capability of real time imaging. However, it is well known that ultrasound(US) images are damaged by the presence of speckles. The main purpose of this thesis is the registration of US prostate images in order to improve the cancer prostate biopsy US assisted procedure. In order to reduce the presence of speckles, a pre-processing over the images taking into account the speckles statistic distribution has been implemented in order to obtain a feature image, which contains the main edges of the prostate US images. With the feature image, a rigid registration using the Normalized Mutual Information is first performed. Then, an elastic registration based on dividing the image into patches and registering rigidly individually those which contain important information has been developed to improve the quality of this first registration. The selection of the patches to be registered has been done using SUSAN corner detector, their registration has been carried out by the Normalized-Cross Correlation, and the connectivity between them has been maintained using a global deformation inversely proportional to the distance between the center of the patches and the pixels of the image. Furthermore, an analysis over all the possible pre-processing options, the main similarity measures, the number of patches and the appropriate corner detector is presented. In the results two main applications of the algorithm developed are presented: the prostate contour tracking and the semi-automatic calibration process. The results show that the tracking process is able to follow the prostate efficiently and if accelerated, it could be used for real time biopsy giving the desirable feedback about the prostate movement and deformation to the physicians.

Time: 17.03.2014, 14:00-14:30
Location: Room 072 in the Fraunhofer IGD, Fraunhoferstrasse 5, S3|05
Speaker: Baran Denis Özdemir (Advisor: Tatiana von Landesberger)
Title: "Visuelle Analyse von Rettungspotenzialen in Ausfalleffekten in Finanznetzwerken" (Bachelor thesis)

Time: 17.03.2014, 14:30-15:00
Location: Room 072 in the Fraunhofer IGD, Fraunhoferstrasse 5, S3|05
Speaker: Kerem Gülensoy (Advisor: Tatiana von Landesberger)
Title: "Visuell-interaktive Exploration und Hypothesen-Bildung von Sentiment- und Aktivitätsdaten" (Master thesis)

GRIS Kolloquium

Time: 12.03.2014, 16:30
Location: Room 074 in the Fraunhofer IGD, Fraunhoferstrasse 5, S3|05
Speaker: Prof. Daniel Scharstein
Department of Computer Science, Middlebury College
Title: "Scalable Surface-Based Stereo Matching"
Abstract: Stereo matching -- establishing correspondences between images taken from nearby viewpoints -- is one of the oldest problems in computer vision. While impressive progress has been made over the last two decades, most current stereo methods do not scale to the high-resolution images taken by today's cameras since they require searching the full space of all possible disparity hypotheses over all pixels. In this talk I will describe a new scalable stereo method that only evaluates a small portion of the search space. The method first generates plane hypotheses from matched sparse features, which are then refined into surface hypotheses using local slanted plane sweeps over a narrow disparity range. Finally, each pixel is assigned to one of the local surface hypotheses. The technique achieves significant speedups over previous algorithms and achieves state-of-the-art accuracy on high-resolution stereo pairs of up to 19 megapixels. I will also present a new dataset of high-resolution stereo pairs with subpixel-accurate ground truth, and provide a brief outlook on the upcoming new version of the Middlebury stereo benchmark.

Bio: Daniel Scharstein, Professor of Computer Science at Middlebury College in Vermont, studied Computer Science at the Universität Karlsruhe, Germany, and received his PhD from Cornell University in 1997. His research interests include computer vision, image-based rendering, and robotics. He maintains several online computer vision benchmarks at He is currently on sabbatical at the German Aerospace Center (DLR) in Oberpfaffenhofen, Germany.

Time: 13.02.2014, 14:00
Location: Room 074 in the Fraunhofer IGD, Fraunhoferstrasse 5, S3|05
Speaker: Dr.-Ing. Hans-Joerg Schulz
University of Rostock
Title: "Towards an Integrated Visual Analysis of Dynamic Networks and Network Dynamics"
Abstract: Time-varying graph data is abundant in today's interlinked world of data networks, power networks, social networks, etc. Its analysis is commonly pursued in two directions: by investigating the dynamic network - i.e., putting the focus on the structural aspects of the data, or by looking at the network dynamics - i.e., putting the focus on the temporal aspects of the data. While the former mainly utilizes analytical methods and visualizations from the domain of graph analysis, the latter applies methods from the domain of time series analysis. This talk aims to show that both of these perspectives are necessary, as each answers to a different set of analysis questions. In the first part of the talk, this is exemplified by introducing two novel methods: a modular degree-of-interest specification for dynamic networks and a transformation approach that produces state-transition-graphs from network dynamics. In the talk's second part, the argument is made for the mutual integration of these two perspectives and a visualization approach is presented that aims to facilitate this integration by means of an in situ embedding of visualizations.

Bio: Hans-Joerg Schulz received his diploma (2004) and doctorate degree (2010) from the University of Rostock, where he is currently working as a postdoctoral researcher in Information Visualization and Visual Analytics. Besides the visualization of dynamic and static graphs, his research interests include visualization design spaces, visualizations of multiple, heterogeneous data sources, and visualization for the biomedical application domain. He maintains the tree visualization survey at and he is a co-guest editor for the 2014 Computers&Graphics Special Section on Visual Analytics. More about his research can be found at

Time: 13.11.2013, 15:30
Location: Room 074 n the Fraunhofer IGD, Fraunhoferstrasse 5, S3|05
Speaker: Prof. Sven Dickinson
Department of Computer Science, University of Toronto
Title: "Perceptual Grouping using Superpixels"
Abstract: Perceptual grouping played a prominent role in support of early object recognition systems, which typically took an input image and a database of shape models and identified which of the models was visible in the image. When the database was large, local features were not sufficiently distinctive to prune down the space of models to a manageable number that could be verified. However, when causally related shape features were grouped, using intermediate-level shape priors, e.g., cotermination, symmetry, and compactness, they formed effective shape indices and allowed databases to grow in size.
In recent years, the recognition (categorization) community has focused on the object detection problem, in which the input image is searched for a specific target object. Since indexing is not required to select the target model, perceptual grouping is not required to construct a discriminative shape index; the existence of a much stronger object-level shape prior precludes the need for a weaker intermediate-level shape prior. As a result, perceptual grouping activity at our major conferences has diminished. However, there are clear signs that the recognition community is moving from appearance back to shape, and from detection back to unexpected object recognition.
Shape-based perceptual grouping will play a critical role in facilitating this transition. But while causally related features must be grouped, they also need to be abstracted before they can be matched to categorical models.
In this talk, I will describe our recent progress on the use of intermediate shape priors in segmenting, grouping, and abstracting shape features. Specifically, I will describe the use of symmetry and non-accidental attachment to detect and group symmetric parts, the use of closure to separate figure from background, and the use of a vocabulary of simple shape models to group and abstract image contours.

Thesis presentations

Time: 04.11.2013, 15:30-16:00
Location: Room 324 in the Fraunhofer IGD, Fraunhoferstrasse 5, S3|05
Speaker: Felix Brodkorb (Advisor: Tatiana von Landesberger)
Title: "Visualisierung von geographisch-basierten Netzwerken" (Master thesis)


Time: 10.10.2013, 10:00-10:30
Location: Room 072 in the Fraunhofer IGD, Fraunhoferstrasse 5, S3|05
Speaker: Zhiping Luo
Title: "Real-time Human Neck Animation Using Simulated Blendshapes"
Abstract: Realistic skinning of a skeletal character is essential to deformable humanoid character animation. Over the years, to achieve that, many efforts are afforded to develop the physical models which practically reproduce life-like skin deformation. Nonetheless, it's well known that the physically-based approaches could be computationally expensive and are only often used in animation films, where the key components can be solved offline and or with a cluster system. To make such physical models applicable in interactive animation and graphics applications, we often have to trade the accuracy for speed. We propose a novel solution to address the trade-off problem. The physical model is developed towards high-fidelity emulation, while only its outcome is used in interactive animation. This technically sound pipeline is experimentally tested for the human neck animation. Because: (i) the complexity of the cervical anatomy makes the physical modeling of the human neck scientifically challenging, (ii) neck animation is important both in facial and head animation. We investigate the biomechanics of the human neck to solve the complex anatomy. Relevant anatomical structures available in a 3D model of human musculoskeletal system are modeled as deformable (based on linear elasticity (continuum) theory) or linked rigid bodies. We couple the soft-hard bodies using soft constraints via elastic springs and form a Lagrangian dynamic system. The simulation of dynamic skin deformation is achieved by automatically binding the vertices to underlying bodies. A blendshape database is generated from the simulation and then used for real-time neck animation. Preliminary results of the muscle simulation and neck animation are provided.

Thesis presentations

Time: 07.10.2013, 10:00-10:30
Location: Room 072 in the Fraunhofer IGD, Fraunhoferstrasse 5, S3|05
Speaker: Johannes Beutel (Advisor: Wissam El Hakimi)
Title: "Sub-Pixel Accurate Fiducial Marker Detection and Pose Estimation from CBCT-Data Date & Location" (Master thesis)
Abstract: Image-guidance plays an important role in modern surgery, like minimally-invasive surgery. It allows high-precision procedures that minimize complications, and increase the efficacy of treatment. A key step in image guided surgery is the accurate registration of preoperatively defined surgical planning data with intra-operative images, acquired using e.g. Cone-beam CT (CBCT). Then it is possible to verify patient position and instrument location. One approach to achieve this alignment is the landmark-based registration. Before acquiring diagnostic data, artificial markers are fixed to the patient's bones, and remain there during the intervention. These fiducial markers are than used to align the intra-operatively acquired 2D-projections to the reference frame of the patient.

Time: 24.09.2013, 11:00-11:30
Location: Room 072 in the Fraunhofer IGD, Fraunhoferstrasse 5, S3|05
Speaker: Jan Lehr (Advisor: Georgios Sakas)
Title: "Segmentation of tomographic data using a symmetrical shape model adaption scheme" (Bachelor thesis)
Abstract: In modern medicine, computer aided diagnostics and surgery planing becomes more and more relevant. Nevertheless, today's machine vision algorithms are not fully reliable with respect to the task of automated organ delineation. To increase segmentation results a bi-directional search is introduced into an existing Active Shape Model segmentation and evaluated on a set of expert segmentations. The experiment's results show, the Active Shape Model segmentation approach can benefit from using a bi-directional search.

Time: 30.07.2013, 17:00-17:30
Location: Room 073 in the Fraunhofer IGD, Fraunhoferstrasse 5, S3|05
Speaker: Benjamin Lück (Advisor: Mate Beljan)
Title: "Kamera-IMU Sensorfusion für mobile Geräte" (Bachelor thesis)
Abstract: (The talk will be given in German.) Für eine Vielzahl von Anwendungen ist es notwendig, die Trajektorie eines in der Hand geführten Smart- phones zu bestimmen. Dafür werden die Daten der im Smartphone eingebauten Sensoren und der Kamera verarbeitet. Die Sensoren werden durch die Kamera unterstützt, da diese sich für die alleinige Bestim- mung der Trajektorie nicht eignen. Die Kamera und die Sensoren werden mittels eines Schachbrettmusters bzw. mit der Allan Varianz Methode kalibriert.
Mit der Inertiale Messeinheit - Inertial Measurement Unit (IMU)- und des Magnetfeldsensors wurde die Lage des Smartphones adäquat bestimmt. Im Anschluss wurde versucht, mittels der kalibrierten Sensoren die Trajektorie zu bestimmen. Dabei trat das Problem auf, dass die Sensoren nicht korrekt kalibiert werden konnten; es ist weiterhin offen.
Die implementierten Computer Vision Algorithmen konnten hingegen die Trajektorie mit der Kamera des Smartphones bestimmen. Zuerst wurde dafür das Verfahren des Feature Matching angewandt, welches sich jedoch als zu langsam herausstellte, um die geforderte Echtzeitfähigkeit auf dem Smartphone zu erreichen. Daher wurde auf die Methode des Feature Trackings zurückgegriffen. Dieses Verfahren ist um Faktor 10 - 20 (abgerundet) schneller und es konnte erreicht werden, dass die Bilder mit ca. 12 Frames per Second (FPS) verarbeitet werden. Die durch die Computer Vision Algorithmen estimierten Trajektorien weisen Abweichungen auf, repräsentieren aber die durchgeführten Bewegungen sehr gut.
Eine weitere offene Fragestellung war, mit welcher Art von Filter die Kamera- mit den Sensordaten
fusioniert werden können. Es wurden grundlegende Filtermechanismen implementiert, welche die Sen- sordaten vereinigen, um die Lage des Smartphone zu bestimmen. Ferner wurde mit einem Kalman Filter die Trajektorie, welche mit der Methode des Feature Tracking bestimmt wurde, verbessert und geglättet.

IREP student project, status report

Time: 30.07.2013, 16:30-17:00
Location: Room 073 in the Fraunhofer IGD, Fraunhoferstrasse 5, S3|05
Speaker: David Ren
University of Illinois at Urbana Champaign
Title: "A Massively Parallel Multimap for GPUs"
(The talk will be given in English.)

GRIS Kolloquium

Time: 10.07.2013, 14:00
Location: Room 074 in the Fraunhofer IGD, Fraunhoferstrasse 5, S3|05
Speaker: Jochen Lang, Ph.D.
University of Ottawa
Title: "Finite Element Based Deformable Object Tracking from a Single Viewpoint"
Abstract: In this talk, I will present our recent work on robustly tracking objects that deform due to an external applied force. The geometry of the object is reconstructed over time based on noisy observations from a single viewpoint. A template mesh of the object in its rest-state is fit to observations in a nonlinear optimization. We use a redundant parameterization of smoothly varying local mesh transformations. While for the observed part of the object, the data term guides the optimization, the unobserved parts of the mesh are only governed by the smoothness term. In a second optimization, we improve the location of the unobserved vertices based on elastic solid deformation solved with finite elements. Synthetic results illustrate the ability of our method to estimate accurate deformations of observed and unobserved parts of an object despite incomplete noisy measurements. We will demonstrate results for deformations recorded by low-quality pointclouds captured either with a commercial stereo camera or a structured light system.
This is joint work with Stefanie Wuhrer, Motahareh Tekieh and Chang Shu.

Time: 03.07.2013, 14:00
Location: Room 074 in the Fraunhofer IGD, Fraunhoferstrasse 5, S3|05
Speaker: Dr. Robert Herzog
Max Planck Institut für Informatik (Saarbrücken)
Title: "Scalable remote rendering and CG-image quality assessment"
Abstract: In my talk I will first present our novel encoding scheme for real-time rendering in the cloud and then introduce a blind image quality metric for detecting rendering-specific artifacts in images.

Remote rendering becomes increasingly attractive for interactive applications since data and application are kept confidential and up-to-date on the server-side and only images are sent to the client, which interacts with the rendering application executed on the server.
However, traditional remote rendering solutions require substantial computational resources on the server. In the first part of my talk, I will present our novel approach, where we shift some computation on the client's GPU to better balance the server/client workload. We tailor our method specifically to 3D content and focus on augmented video information (depth and motion vectors). This allows us to perform efficient video upsampling on the client while reducing the workload on the server by rendering and encoding lower resolution images.

On the other hand, generating synthetic images of complex 3D scenes using photo-realistic rendering software is also prone to artifacts and requires expert knowledge to tune the parameters. Although the manual work required for detecting and preventing such artifacts can be automated through objective quality evaluation, most methods rely on a ground-truth reference, which is often not available in rendering applications. Hence, in the second part of my talk, I will present a no-reference (blind) image quality metric dedicated to rendering artifacts whose performance is on par with state-of-the-art metrics that do require a reference. This level of predictive power is achieved by exploiting information about the underlying synthetic scene (e.g., 3D surfaces, textures) instead of merely considering color, and training our image quality metric for typical rendering artifacts in a supervised fashion.

Time: 17.06.2013, 11:30
Location: Room 073 in the Fraunhofer IGD, Fraunhoferstrasse 5, S3|05
Speaker: Prof. Kwangyun Wohn
KAIST, Korea
Title: "Culture Technology: Beyond Technology, Beyond Culture"
Abstract: Around 1994, I coined Culture Technology (CT) with my personal perspective of setting a framework for the symbiosis between technology and culture, more specifically between the digital technology and cultural artifacts. Somehow, the terminology, CT, has become popular in Korea, and for some, mysterious and political reasons, has been adopted by the Korean government as one of the Growth Engine for the national advancement. There have been many good, bad, and ugly things happened since then, but one of the excitements was the establishment of a new graduate program at KAIST, primary objective of which was to facilitate the interdisciplinary studies in digital technology and arts-culture.
In this talk, I will begin with the overview of this graduate school, GSCT, including its role, major accomplishments, and some of the noteworthy research activities. I then will switch the topic - my own research on virtual environments, mostly on realtime rendering, displaying, and interactions. I will conclude my presentation with my personal perspective on the future of CT and its implication to virtual reality.

Bio: Kwangyun Wohn is Professor and Founding Dean of the GSCT (Graduate School of Culture Technology) at KAIST (Korea Institute of Science and Technology). Before he joined at KAIST, he had been with Harvard University as Lecturer, and with University of Pennsylvania as Assistant Professor. Having returned back to his home country, Korea, he had been with Computer Science Department for fifteen years, until he established a new graduate school, Graduate School of Culture Technology in 2005.
Major activities and accomplishments include: Director of Virtual Reality Research Center which is a national center of research excellence, Founding President of Korean Society of Human-Computer Interaction, Founding President of Korean Society of Performing Art, and Editorial Board of British Computer Society. While his research interest spans the broad range of the intersection between art and science – from theoretical aspects to practicalities – he focuses his research efforts to the application of virtual reality technology to various cultural artifacts such as stage performances, museum exhibitions and entertainment contents.

Thesis presentations

Time: 12.06.2013, 10:30-11:00
Location: Room 5/6 in the TU Darmstadt Instituts- und Verwaltungsgebäude, Rundeturmstraße 12, S3|19
Speaker: Nicolas Weber (Advisor: Sven Widmer)
Title: "Construction of Ray-Tracing Acceleration Structures in an Out-of-Core Multi-GPU Environment" (Master thesis)
Abstract: The construction of ray-tracing acceleration structures is an important factor in ray-tracing today. Though GPUs have shown to provide massive computational power, their usage for constructing acceleration structures is limited as the available memory is too small for vast 3D scenes. Construction algorithms that have been used so far lack the necessary parallelism in the first tree levels. We propose a new out-of-core construction algorithm for constructing a binned SAH based KD-Tree for arbitrarily big scenes on a multi-GPU system, which overcomes the problem of the limited memory while also utilizing parallelism even at the first tree level.

Time: 12.06.2013, 11:15-11:45
Location: Room 5/6 in the TU Darmstadt Instituts- und Verwaltungsgebäude, Rundeturmstraße 12, S3|19
Speaker: Carsten Haubold (Advisor: Sven Widmer)
Title: "Out-of-core Bidirectional Path-Tracing on a Multi-GPU System" (Master thesis)
Abstract: Rendering photorealistic images of small to moderate scenes using GPU path tracing has become a widely used technique. Now artists and architects strive to create larger and more detailed models, but the sheer amount of data makes path tracing on GPUs infeasible. The problems range from data exceeding available GPU memory to low GPU occupancy as incoherent rays block executions on SIMT architectures.

This thesis presents a framework for out-of-core processing on a multi-GPU system, which automatically performs on-demand transfers of memory between host and devices, and allows for flexible work scheduling strategies. Based on this framework we introduce an unbiased Bidirectional Path-Tracer, which is designed to work on larger scenes than conventional approaches. The scene data is assumed to be present in a tree acceleration structure split up into treelets containing geometry and material data which t into GPU memory. We perform ray tracing in batches for single treelets, sorting and rescheduling the rays for other treelets which they cross along their path through the scene. We show that queuing rays per treelet for coherence before tracing significantly increases performance, and compare the benefits of different treelet construction algorithms.

GRIS Kolloquium

Time: 16.05.2013, 11:00
Location: Room 074 in the Fraunhofer IGD, Fraunhoferstrasse 5, S3|05
Speaker: Ivan Laptev, Ph.D.
Title: "Human action recognition: recent progress, open questions and future challenges"
Abstract: Action recognition is a active research area with many potential applications. In this talk I will briefly review the past of action recognition and will describe our recent efforts in representing and learning actions in video. I will then change the viewpoint and will discuss the problem definition as well as the relation of action recognition to other visual recognition tasks. I will argue that our future algorithms, to be more practical, should probably be more person-centric and should analyze scenes and objects in terms of their function, i.e. what people can do and typically do with them. I will conclude with some of our recent efforts in this direction.
Bio: Ivan Laptev is a full-time researcher in the WILLOW team at INRIA Paris and Ecole Normale Superieure. He has received his PhD in Computer Science from the Royal Institute of Technology (KTH) in 2004 and his Master of Science degree from the same institute in 1997. He has been a research assistant at the Technical University of Munich (TUM) during 1997-1999 and he has joined INRIA in 2004. Ivan's main research interests concern visual understanding of dynamic scenes including recognition of human actions, scenes, objects and interactions. Ivan has published over 40 papers at international conferences and journals on computer vision, he serves as an associate editor of International Journal of Computer Vision and Image and Vision Computing Journal, he was an area chair for CVPR 2010, ICCV 2011, ECCV 2012 and CVPR 2013. He is a winner of the ERC junior grant in 2012.

Thesis presentations

Time: 03.05.2013, 10:00-10:30
Location: Room 073 in the Fraunhofer IGD, Fraunhoferstrasse 5, S3|05
Speaker: Daniel Thürck (Advisor: Arjan Kuijper)
Title: "A well-posed parameter-free model for nonlinear diffusion and its applications in mobile image processing" (Bachelor thesis)
Abstract: Images and videos today represent our most important media. Recently, taking pictures and recording videos with mobile phones and uploading them to the internet has become common, especially in social networks. However, due to low-quality CCD sensors, those pictures often suffer from noise. A solution here would be to use quite well-known image processing algorithms, especially anisotropic diffusion. The most famous model, the Perona and Malik equation, unfortunately is ill-posed and thus is problematic.
In this work, we present an alternative model for anisotropic diffusion that is constructed in a bottom-up fashion for denoising and well-posedness. The problem of setting the matching input parameters for denoising is tackled by the use of machine learning techniques. Ultimately, we present a prototypical implementation on embedded hardware that shows that the use of such sophisticated techniques is possible for mobile use.

Time: 29.04.2013, 14:00-14:30
Location: Room 072 in the Fraunhofer IGD, Fraunhoferstrasse 5, S3|05
Speaker: Martin Hess (Advisor: Tatiana von Landesberger)
Title: "Interactive visual Comparison of many hierarchic Datasets" (Master thesis)
Abstract: Hierarchical Structures are widely used for multiple purposes. One example is the current research in phylogenetics, especially the comparison of large numbers of so called phylogenetic trees. The main target is the discovery of similarities and dissimilarities between those trees. The size, complexity and large amount of those trees makes the comparison task very difficult for the user.
This master-thesis presents a new approach for the visual and interactive comparison of a large set of trees, considering their elements, topology and the parameters used for their calculation.
By using these concepts, the user is able to compare more than 1000 trees at once. The new method is applied to the comparison of 1344 different phylogenetic trees of the 16S ribosomal RNA from different bacteria.

Time: 26.03.2013, 16:30-17:00
Location: Room 074 in the Fraunhofer IGD, Fraunhoferstrasse 5, S3|05
Speaker: Andreas Franek (Advisor: Jens Ackermann)
Title: "Enhancing GPS Precision using Structure from Motion" (Master thesis)
Abstract: GPS positional data has been used to enhance Structure from Motion (SfM) reconstructions of large scenes [Lhu11, Ell04]. Here we evaluate the opposite approach; utilizing SfM reconstructions of scenes between 50 and 500m to enhance the GPS meta data associated with the images used in SfM. These images are here taken from online community photo collections. The statistical properties of SfM reconstructed camera positions and GPS measurements are considered and found to be very distinct. Further, the distribution of geotags created manually by users and that of GPS receiver created geotags in online community photo collections is examined. The result shows that images in online community photo collections scarcely contain reliable GPS meta data.
Because of this it is hard to estimate a transformation between GPS and SfM space. With such a transformation it is possible to improve the accuracy of inaccurate geotags and create geotags for images not containing one previously. An algorithm is designed that suits the statistical properties of SfM reconstructions and GPS measurements.
Using this algorithm it is possible to enhance the geotags of SfM image bundles even when only few reliable are present initially.

GRIS Kolloquium

Time: 19.02.2013, 14:00
Location: Room 074 in the Fraunhofer IGD, Fraunhoferstrasse 5, S3|05
Speaker: Roy Ruddle
Universitaet Leeds, England
Title: "The Leeds Virtual Microscope"
Abstract: The Leeds Virtual Microscope is an interactive visualization system, capable of rendering gigapixel virtual slides onto high-resolution, wall-sized displays. I will describe the six years of research we have conducted, developing the Microscope, and evaluating its use for the diagnosis of cancer, the training of histopathology specialists, and the education of medical students.

For further information, see

Time: 24.01.2013, 12:00
Location: Room 073 in theFraunhofer IGD, Fraunhoferstrasse 5, S3|05
Speaker: Engin Deniz Diktas
Title: "Rendering and Acceleration Structures"
Abstract: This talk gives a broad overview over my research in computer graphics and its applications. The talk includes topics such as acceleration structures for haptic rendering, collision detection and distance computations and their benchmarking; advanced rendering methods like Compressed ID-Shadow Maps, ID-based Silhouette Rendering, Ray Tracing on the GPU, Rain and Snow effects as well as a custom Anti-Aliasing method for rendering.

Thesis presentations

Time: 22.01.2013, 16:30-17:00
Location: Room 074 in the Fraunhofer IGD, Fraunhoferstrasse 5, S3|05
Speaker: Oliver Goroll (Advisor: Dominik Wodniok)
Title: "Multi-View Stereo via Semi-Global Matching" (Bachelor thesis)
Abstract: (The talk will be given in German.) In dieser Arbeit wird ein Ansatz vorgestellt, wie mit adäquaten Kostenfunktionen und einem effizientem Matching-Algorithmus Tiefenkarten erstellt werden, die zu einer guten Rekonstruktion verwendet werden. Zur Kostenberechnung werden die Census Transform und eine hierarchische Version von Mutual Information zusammen mit Semi-Global Matching verwendet. Die entstehenden Tiefenkarten zu einem Base-Bild und mehreren Match-Bildern werden kombiniert, um die Fehlerzahl zu verringern und die Abdeckung zu erhöhen. Der beschriebene Ansatz wird durch eine prototypische Implementierung anhand verschiedener Testdaten evaluiert.

Time: 22.01.2013, 17:15-17:45
Location: Room 074 in the Fraunhofer IGD, Fraunhoferstrasse 5, S3|05
Speaker: Fabian Langguth (Advisor: Dominik Wodniok)
Title: "Guided Capturing of Multi-view Stereo Datasets" (Master thesis)
Abstract: We present an application for mobile devices, that allows any user, even without background in computer vision, to capture a complete set of images, that is suitable for a multi-view reconstruction. Compared to related tasks, such as panorama capture, this setting is much harder, as the camera needs to move unrestricted in 3D space. Our system uses structure from motion to register captured images and generates a sparse reconstruction of the scene. The dataset is built in an incremental procedure, where the next best view is computed with a novel view planning strategy, that aims for a good coverage of the scene. The user is then guided towards the new view, and the image is captured automatically at the right position. The next iteration starts after the reconstruction has been updated. The quality of the resulting dataset is on par with datasets captured by an expert user.

Time: 17.12.2012, 14:00-14:30
Location: Room 324 in the Fraunhofer IGD, Fraunhoferstrasse 5, S3|05
Speaker: Sebastian Fahnenschreiber (Advisor: Tatiana von Landesberger)
Title: "Graph layout based on MEU" (Bachelor thesis)
(The talk will be given in German.)

Time: 14.12.2012, 10:00-10:30
Location: Room 220 in the Fraunhofer IGD, Fraunhoferstrasse 5, S3|05
Speaker: Hendrik Lücke-Tieke (Advisor: Thorsten May)
Title: "Stabiles Graph-Layout basierend auf lokalen Layouts" (Bachelor thesis)
Abstract: (The talk will be given in German.) Die Effektivität einer visuellen Analyse mit Hilfe von Node-Link-Diagrammen wird durch 2 Faktoren stark beeinflusst: die Nachvollziehbarkeit von Zusammenhängen und die Übersichtlichkeit der Visualisierung. Viele Visualisierungsverfahren für Node-Link-Diagramme betrachten den gesamten Datensatz und erzeugen daraus eine Gesamtdarstellung, ein sogenanntes globales Layout. Dieses globale Layout erleichtert dem Betrachter die Orientierung, denn jeder Datenpunkt befindet sich immer an der gleichen Position. Ein Analyst kann sich so auf den Lageplan in seinem Kopf verlassen, die Visualisierung ist in diesem Sinne stabil. Dies ist allerdings ein Kompromiss auf Kosten der Genauigkeit lokaler Zusammenhänge. Doch genau diese Zusammenhänge können relevant sein und bedürfen daher einer gesonderten Betrachtung ohne Abhängigkeiten oder Einflüsse des restlichen Graphen. Ein solcher, unabhängig berechneter Ausschnitt wird als lokales Layout bezeichnet.
Die statische, interaktive Exploration eines Graphen wird durch Filterung des globalen Layouts realisiert. Die dynamische, interaktive Exploration eines Graphen basiert dagegen auf einem modifizierten lokalen Layout. Mit jeder Änderung des darzustellenden Ausschnitts wird ein neues lokales Layout berechnet, welches dem vorherigen Layout ähnelt. Mit diesem Verfahren können lokale Zusammenhänge übersichtlich und nachvollziehbar visualisiert werden. Aber unterschiedliche Explorationsverläufe resultieren in unterschiedlichen Layouts, der Transfer des Wissensaus einer Exploration auf eine andere Exploration ist schwierig. Gibt es also eine Methode, die die Vorteile beider Verfahren verbinden kann? Eine Methode, die es dem Anwender erlaubt, sich auf seine Orientierung zu verlassen und trotzdem lokale Zusammenhänge des betrachteten Ausschnitts übersichtlich darzustellen? Dieser Frage widmet sich die vorliegende Arbeit.

GRIS Kolloquium

Time: 13.12.2012, 16:30
Location: Room 074 in the Fraunhofer IGD, Fraunhoferstrasse 5, S3|05
Speaker: Carsten Rother, Ph.D.
Computer Vision Group (Microsoft Research Cambridge, UK)
Title: "From Particle Stereo to Scene Stereo"
Abstract: In this talk I will present two lines of research which are both applied to the problem of stereo matching. The first line of research tries to make progress on the very traditional problem of stereo matching. Last year I presented in Darmstadt the PatchmatchStereo work which gets surprisingly good results with a simple energy function consisting of unary terms only and using for optimization the well-known PatchMatch method, which was designed for image editing purposes. This year we extended this work by adding to the energy function the standard pairwise smoothness terms. The main contribution of this work is the optimization technique, which we call PatchMatch-BeliefPropagation (PMBP). It is a special case of max-product Particle Belief Propagation, with a new sampling schema motivated by Patchmatch. The method may be suitable for many energy minimization problems in computer vision, which have a non-convex, continuous and potentially high-dimensional label space. The second line of research combines the problem of stereo matching with the problem of object extracting in the scene. We show that both tasks can be solved jointly and boost the performance of each individual task. In particular, stereo matching improves since objects have to obey physical properties, e.g. they are not allowed to fly in the air. Object extracting improves, as expected, since we have additional information about depth in the scene.

Bio: Carsten Rother received the Diploma degree with distinction in 1999 from the University of Karlsruhe. He did his PhD at the Royal Institute of Technology Stockholm, supervised by Stefan Carlsson and Jan-Olof Eklundh. His thesis was nominated for the Best Nordic Thesis Award 2003-2004, as one out of two candidates from Sweden. From 2003-04, he was PostDoc at Microsoft Research Cambridge, and since 2004 has been a permanent researcher there. His research interests are in the field of “(Markov) Random Field Models for Computer Vision”, “Discrete Optimization”, and “Vision for Graphics” (in particular interactive segmentation and matting). He has published more than 20 high impact papers (at least 10 citations) at international conferences and journals, and co-authored a paper which won the best paper honourable mention award at CVPR’05. He serves on the program committee of major conferences (e.g. SIGGRAPH, ICCV, ECCV, CVPR, NIPS), and has been area chair for BMVC‘08, and ‘09. He has organized some workshops (“Interactive computer vision” at ICCV’07, “Color and Reflectance” at ICCV’09), and supervises several PhD students.

Thesis presentations

Time: 07.11.2012, 16:30-17:00
Location: Room 074 in the Fraunhofer IGD, Fraunhoferstrasse 5, S3|05
Speaker: Sebastian Koch (Advisor: Dominik Wodniok)
Title: "Development of a Mobile Projector Camera System for Structured Light Scanning" (Master thesis)
Abstract: The aim of this thesis is to perform geometric self calibration of a projector camera system using only a smartphone and a small projector. The developed system is capable of performing Structured Light Scanning with off-the-shelf devices, namely a Nokia N900 smartphone and an Optoma PK 320 pico projector. We decouple the projector from the smartphone, which is our computational device and our camera. No information needs to be transferred through a direct connection between the projector and the smartphone during capture and digitisation. The system is capable of reconstructing rigid objects while the projector is being held in one hand and the camera is fixed. It can reconstruct dense depth maps of objects ranging in size from 5 to 50 cm. The performance of the system using three different types of pattern codification strategies is evaluated in terms of calibration accuracy and reconstruction quality. A system is presented capable of capturing the geometry of an object in less than one second and reconstructing it in less than 30 minutes on the smartphone or in less than 30 seconds on a current desktop PC.

Time: 07.11.2012, 17:15-17:45
Location: Room 074 in the Fraunhofer IGD, Fraunhoferstrasse 5, S3|05
Speaker: (Advisor: Dominik Wodniok)
Title: "Improving Cache usage of Tracing Incoherent Rays on GPUs" (Diploma thesis)
Abstract: Path tracing and related global illumination techniques create beautiful photorealistic images but computing these images is expensive. The process can be sped up by parallelization as it is embarassingly parallel. With recent developments in CPU technology moving towards many-core architectures and GPUs becoming more general purpose architectures, path tracing can now be parallelized on commodity hardware. Unfortunately, while the parallelization is trivial in theory, in reality hardware details make it more difficult, especially for tracing incoherent rays. This thesis investigates the impact of different bounding volume hierarchy (BVH) and node memory layouts as well as accessing the BVH in different memory areas on the ray tracing performance of a path tracer on a many-core wide SIMD architecture by NVIDIA, the Tesla C2070. Furthermore, we optimize the BVH layout by using information gathered in a pre-processing pass which we use in a number of different BVH reordering techniques. Depending on the memory area and complexity of the scene, we are able to achieve a speedup ranging from negligible to moderate.

GRIS Kolloquium

Time: 06.11.2012, 16:30
Location: Room 074 in the Fraunhofer IGD, Fraunhoferstrasse 5, S3|05
Speaker: Johannes Kopf, Ph.D.
Microsoft Research, Redmond
Title: "Recent Work on Digital Reconstruction of Comic Art and Predictive High Quality Image Completion"
Abstract: In this talk I will present two recent projects from our lab that are to be published at SIGGRAPH Asia 2012. In the first part I will examine the digitization of printed color illustrations, and in particular, comic art. Over the last few decades, nearly all media (e.g., audio, images, video, text, CAD) have been transitioning to digital representations. There are of course countless works that predate this transition. One key challenge is that analog representations in physical media may be subject not only to noise but to other artifacts specific to each physical representation. I will present our recent work on automated conversion of scanned color comic books, as well as some other types of hand-drawn color illustrations, into a new, high-fidelity scalable digital representation, suitable for today’s high resolution digital reading devices. Our approach is to model the color comics printing process in a rigorous manner and to invert this model using non-linear optimization. In the second part of the talk I will present novel work on high quality image completion. Our method uses automatically derived search space constraints, which lead to improved texture synthesis and semantically more plausible results. The constraints also facilitate performance prediction before the completion is produced. We use our predictive ability to automatically crop and then complete stitched panoramas. Our optimized crops include as much of the original panorama as possible while avoiding regions that can be less successfully filled in.

Time: 01.11.2012, 13:00
Location: Room 074 in the Fraunhofer IGD, Fraunhoferstrasse 5, S3|05
Speaker: Andreas Geiger
Karlsruhe Institute of Technology (KIT)
Title: "Inspection of Complex Objects using Multiple X-ray Views"
Abstract: Navigating a car safely through complex environments is considered a relatively easy task for humans. Computer algorithms, however, can't nearly match human performance and often rely on 3D laser scanners or detailed maps. The reason for this is that the level and accuracy of current computer vision and scene understanding algorithms is still far from that of a human being. In this talk I will argue that pushing these limits requires solving a set of core computer vision problems, ranging from low-level tasks (stereo, optical flow) to high-level problems (object detection, 3D scene understanding). First, I will introduce the KITTI datasets and benchmarks with accurate ground truth for evaluating stereo, optical flow, SLAM and 3D object detection/tracking on realistic video sequences. Results from state-of-the-art algorithms reveal that methods ranking high on established datasets such as Middlebury perform below average when being moved outside the laboratory to the real world. Second, I will propose a novel generative model for 3D scene understanding that is able to reason jointly about the scene layout (topology and geometry of streets) as well as the location and orientation of objects. By using context from this model, performance of state-of-the-art object detectors in terms of estimating object orientation can be significantly increased. Finally, I will give an outlook on how prior information in form of large-scale community-driven maps (OpenStreetMap) can be used in the context of 3D scene understanding.
Bio: Andreas Geiger studied computer science and mathematics at Karlsruhe Institute of Technology (KIT) in Germany. During his studies he spent 6 months at EPFL (working with Pascal Fua and Vincent Lepetit) and 6 months at MIT (working with Raquel Urtasun and Trevor Darrell). Currently he is a 4th year doctoral candidate at KIT under the supervision of Christoph Stiller (KIT) and Raquel Urtasun (TTI-C). His research interests are in computer vision and machine learning, with a focus on vision-based 3D scene understanding.

Thesis presentations

Time: 29.10.2012, 10:30-11:00
Location: Room 073 in the Fraunhofer IGD, Fraunhoferstrasse 5, S3|05
Speaker: Sebastian Freutel (Advisor: Matthias Kirschner)
Title: "Formmodellbasierte Segmentierung der Leber in kontrastverstärkten CT-Scans" (Master thesis)
Abstract: (The talk will be given in German.) Die vorliegende Masterarbeit beschäftigt sich mit der automatischen Segmentierung der Leber in kontrastverstärkten CT-Aufnahmen. Ein bestehender Active-Shape-Model-Algorithmus wurde durch neue Methoden erweitert, um die Segmentierungsleistung zu verbessern. Dabei wurde ein zusätzlicher Energieterm eingesetzt, um eine glatte Oberfläche in der Segmentierung zu erzwingen. Außerdem wurden neue Fitnessfunktionen für das Erscheinungsmodell entworfen. Des Weiteren wurden neue Methoden der Normalisierung der Bilddaten entwickelt, sowie Methoden zur robusten Schätzung der Intensitätsverteilung der Leber in den Bilddaten ausgearbeitet.
Die neuen Ansätze wurden in einer Vielzahl von Experimenten evaluiert, sie konnten die Segmentierung teilweise erheblich verbessern. Es wurde festgestellt, dass eine Aufteilung der in den Fitnessfunktionen verwendeten Intensitätsprofile in einen inneren und äußeren Teil gute Ergebnisse erzielt, ebenso dieselbe Aufteilung bei der Normalisierung. Außerdem erwies sich ein partiell globales Erscheinungsmodell als besser geeignet als ein vollständig lokales Modell. Die vorgeschlagenen Energieterme sorgten ebenfalls für eine deutlich höhere Qualität der Segmentierung, die Reduzierung der Anzahl der dafür benötigten Parameter erfordert jedoch weitere Arbeiten. Die neu entwickelten Methoden der Intensitätsschätzung verringerten den Segmentierungsfehler in den meisten Fällen. Die Experimente legen nahe, dass die Ergebnisse durch eine größere Anzahl Trainingsdaten noch deutlich verbessert werden können.

Time: 29.10.2012, 14:00-14:30
Location: Room 073 in the Fraunhofer IGD, Fraunhoferstrasse 5, S3|05
Speaker: Olav Lenz (Advisor: Tatiana von Landesberger)
Title: "Visual Analysis of Node Importance and Patterns in Network Contagions" (Master thesis)

Time: 19.10.2012, 14:00-14:30
Location: Room 073 in the Fraunhofer IGD, Fraunhoferstrasse 5, S3|05
Speaker: Richard Gross (Advisor: Arjan Kuijper)
Title: "Multimodal Kinect-supported Interaction for the Visually Impaired" (Master thesis)
Abstract: This thesis suggests a new computer interface, specifically targeted at blind and visually impaired people. We use the Microsoft Kinect to track a users position and have implemented a novel spatial interface to control text-to-speech synthesis of a document. Which actions are executed is solely determined through hand movements in relation to the body. All feedback for the actions is given in auditory form, through synthesized speech or earcons. Earcons are brief, unique sounds that convey information. Visually impaired or blind users do not have to point or remember keyboard commands, but can use their proprioceptive sense to effectively explore documents and execute actions.
The test results are encouraging. Even when participants found themselves lost they were always able to find their way back to an interface state they knew how to navigate. Furthermore, most negative feedback can be attributed to the current technical limitations and not the spatial interface itself.

GRIS Kolloquium

Time: 05.10.2012, 11:00
Location: Room 073 in the Fraunhofer IGD, Fraunhoferstrasse 5, S3|05
Speaker: Prof. Domingo Mery
Universidad Católica de Chile, Santiago de Chile
Title: "Inspection of Complex Objects using Multiple X-ray Views"
Abstract: This talk presents a new methodology to identify parts of interest inside a complex object using multiple X-ray views. The proposed method consists of three steps: i) Structure estimation: It obtains a geometric model of the multiple views of the object under test. The geometric model is estimated by a bundle adjustment algorithm on stable SIFT key points across the multiple views which are not necessary to be sorted. ii) Parts detection: It detects parts of interest of the object using an ad-hoc segmentation algorithm (application dependent) followed by a tracking algorithm based on geometric and appearance constraints. During segmentation algorithm, false alarms are allowed, because they will be eliminated by tracking algorithm without discriminating the parts of interest. iii) Analysis: It makes the final decision by examining tracked objects in 2D, from different views, and in 3D, from a reconstructed object using a local binary computed tomography algorithm. In order to illustrate the effectiveness of the proposed method, the algorithm was tested on several sequences with 1 to 9 views in five applications using 2D analysis. Additionally, we present preliminary results of 3D analysis for semi-automatic baggage screening.

Thesis presentations

Time: 24.09.2012, 14:00-14:30
Location: Room 048 in the Fraunhofer IGD, Fraunhoferstrasse 5, S3|05
Speaker: Dennis Gehrke (Advisor Matthias Kirschner)
Title: "Konsistente und verzerrungsminimierende Parametrisierung von Oberflächen sphärischer Topologie" (Diploma theses)
Abstract: (Talk will be given in German) In dieser Arbeit wird gezeigt, wie punktweise Korrespondenzen zwischen ähnlichen, polygonal repräsentierten Objekten hergestellt werden können. Bei diesen Objekten handelt es sich um Genus-0-Dreiecksnetze von biologischen Organen, die durch bildgebende Verfahren gewonnen wurden. Dabei ist es nicht erforderlich, manuell eine Menge von korrespondierenden Punktpaaren zwischen den Objekten zu definieren. Der vorgestellte Algorithmus ermittelt vielmehr automatisch Korrespondenzabbildungen zwischen allen beteiligten Objekten.

Für den Korrespondenzaufbau werden die beteiligten Objekte auf eine Kugeloberfläche, die als gemeinsame Basisdomäne dient, abgebildet. Dazu werden die Meshes zunächst zu einem Tetraeder vereinfacht, der trivial auf eine Sphäre abgebildet werden kann. Dann wird der Tetraeder schrittweise wieder zu seiner ursprünglichen Form verfeinert. In jedem Verfeinerungsschritt wird für den hinzugekommenen Knoten eine geeignete Position auf der Kugeloberfläche gesucht.

Zur Vereinfachung und anschließenden Verfeinerung wurde zunächst eine generische Progressive- Mesh-Implementierung in ein bestehendes Softwaresystem zur Segmentierung anhand statistischer Formmodelle hinzugefügt. Die Metrik zur Vereinfachung des Meshes ist austauschbar und diverse Metriken wurden implementiert und verglichen. Die Implementierung des Progressive- Mesh-Aufbaus stellt dabei jedoch immer sicher, dass das Mesh seine topologischen Eigenschaften erhält. Diese Progressive-Mesh-Darstellung liefert den initial einzubettenden Tetraeder sowie die Reihenfolge der Verfeinerungsschritte.

Das Vorgehen zum Finden einer geeigneten Position der Knoten auf der Sphäre in den Verfeinerungsschritten muss in erster Linie sicherstellen, dass sich das Dreiecksnetz korrekt in die Sphäre einbettet, d. h. sich keine Kanten kreuzen und Dreiecke damit überlappen. Ein zweiter Aspekt ist das Finden einer Position, welche die zwangsläufig durch die Einbettung auf die Sphäre erzeugte Flächen- und Winkelverzerrung minimiert.

So gewonnene Korrespondenzabbildungen haben eine Vielzahl von Anwendungen. Der primäre Fokus liegt hier im anschließenden Aufbau von statistischen Formmodellen, welche dann zur Segmentierung von gleichen Organen in medizinischen Daten verwendet werden können.


Time: 16.08.2012, 14:00-14:30
Location: Room 072 in the Fraunhofer IGD, Fraunhoferstrasse 5, S3|05
Speaker: Anne-Sophie Ettl (Advisor Arjan Kuijper)
Title: "Klassifikation von Bildbereichen in digitalisierten Dokumenten zur Anwendung auf mobilen Geräten" (Bachelor theses)
Abstract: (Talk will be given in German) Die weiterführende Verarbeitung und maschinelle Interpretation von Bildmaterial spielt eine immer wichtigere Rolle im Mobile-Bereich. Sowohl für die effiziente Kompression als auch für die maschinelle Auswertung von Texten bietet es sich an, Textinhalte nicht als Grafik, sondern als Textinformation zu speichern. Hierzu soll mit der Kamera eines Smartphones aufgenommenes Bildmaterial von Magazinen nach Text- und Bildbereichen klassifiziert werden. In der Arbeit werden etablierte Desktop-Verfahren vorgestellt und hinsichtlich ihrer Anwendungmöglichkeiten auf Mobilgeräten untersucht. Im Anschluss wird ein Ansatz für ein Verfahren zur Bildsegmentierung entwickelt, das die begrenzten Ressourcen der mobilen Geräte berücksichtigt. Eine prototypische Implementierung wird in Python mithilfe der OpenCV-Bibliothek realisiert.

Time: 13.08.2012, 15:00-15:30
Location: Room 073 in the Fraunhofer IGD, Fraunhoferstrasse 5, S3|05
Speaker: Florian Jung (Advisor Matthias Kirschner)
Title: "Automatische Detektion von Organen in CT-Bildern auf Basis des Viola-Jones-Verfahrens" (Master theses)
Abstract: (Talk will be given in German) Die Segmentierung von Organen spielt in der medizinischen Bildverarbeitung eine entscheidende Rolle und ist Voraussetzung für eine Vielzahl von Problemstellungen. Die gängigen Segmentierungsverfahren arbeiten alle lokal, was zur Folge hat, dass zuerst eine Positionsbestimmung für das gesuchte Organ durchgeführt werden muss. Für diese Organdetektion auf Computertomographiebildern wurden bereits eine Vielzahl an Verfahren vorgestellt. Die Mehrheit dieser Ansätze arbeitet jedoch anwendungsspezifisch und verwendet a priori-Wissen, weshalb diese Verfahren nur zu der Detektion von genau einem Organ eingesetzt werden können. Wir stellen einen generischen Ansatz vor, der lernbasiert arbeitet und somit komplett ohne Vorwissen auskommt. Dadurch sind wir in der Lage, einen Detektor für beliebige Organe zu trainieren, der sich durch seine Detektionsgeschwindigkeit und -genauigkeit auszeichnet. Als Grundlage für den von uns entwickelten Algorithmus dient das Gesichtserkennungsverfahren von Viola und Jones, das wir ins Dreidimensionale portiert haben. Zusätzlich haben wir einige Modifikationen vorgenommen und das Verfahren um bedeutende Funktionalität erweitert, um die Qualität der Ergebnisse weiter zu optimieren, die Ausführungszeit zu verkürzen und den Algorithmus robuster zu gestalten.

GRIS Kolloquium

Time: 12.07.2012, 14:00 Uhr
Location: Raum 074 im Fraunhofer IGD, Fraunhoferstrasse 5, S3|05
Speaker: Prof. Jose M. Pena
CeSViMa, Universidad Politecnica de Madrid, Spain
Title: "Neuroscience Applications: Soft Computing and Simulation Techniques"
Abstract: Computational neuroscience is a challenging field in which multiple disciplines collaborate in the improvement of our knowledge of the brain and its function. The main aspect in which this field differentiates from experimental neuroscience is in the use of computational-based techniques, to either gather or interpret experimental data. This talk presents the active research line in the use of soft computing techniques (data mining, machine learning and optimization techniques) in their different applied contributions: -Combined soft computing and image processing techniques in the segmentation and labeling of microscopy data (for both light and electron microscopy). -Multi-scale simulation: (i) molecular dynamics to signal transmission models, (ii) neurite growing simulation, and (iii) mechanical-electrophysiological coupling, -Interactive visual analytics and exploratory data navigation (for either neuroanatomical or clinical information).

Bio: Prof. José María Peña is professor at the UPM and scientific subdirector of the CeSViMa-UPM Supercomputing Center. He has PhD in Computer Science from the UPM. Prof. Peña has developed his career in the field of high-performance data analysis and modeling and in the last year an active research in applied problems in the domain of genomics, proteomics and neuroscience. He has more than 15 years of research experience backed-up by participation in more than 20 national and international projects, leading 5 of them. Additionally, he is actually the coordinator of the UPM node in two major computational neuroscience initiatives: Cajal Blue Brain project (Spanish participation in the international Blue Brain Project) and Alzheimer 3 (a private-funded initiative in new methods in the research and treatment of Alzheimer’s disease), in addition Prof. Pena has a major role in the coordination of the Neuroinformatics activities within the Human Brain Project (HBP) FET-Flagship pilot project. He is member of the Intelligent Data Analysis (IDA) Council and associate editor of several journals. In 2006, he received with the Best Young Researcher Award from FGUPM for researchers under 35 years old. He has published more than 100 peer-reviewed contributions (in international journals and conferences).

Time: 29.05.2012, 14:00
Location: Room 073 in the Fraunhofer IGD, Fraunhoferstrasse 5, S3|05
Speaker: Prof. Jorg Peters
Dept. of CISE, University of Florida
Title: "Pixel-accurate Display of Spline Patches"
Abstract: How finely should one evaluate a spline patch to see no artifacts on the screen? How can one characterize artifacts in the first place?

This talk answers both questions and gives an algorithm for interactive display using the latest GPU support. This is illustrated by an example of interactive movie generation.

Bio: Dr. Jorg Peters is Professor of Computer and Information Sciences at University of Florida. He is interested in representing, analyzing and computing with geometry. To this end he has developed new tools for free-form modeling and design in spline, Bezier, subdivision and implicit representations. He is heading the TIPS project enabling surgeon-educators to author VR-based simulations with force feedback.

Dr. Peters obtained his Ph.D. in 1990 in Computer Sciences from the University of Wisconsin, with Carl de Boor as advisor. In 1991 and 1992, Dr Peters held positions at the IBM T. J. Watson Research Center and Rensselaer Polytechnic Institute before moving to the computer science department of Purdue University. In 1994, Dr. Peters received a National Young Investigator Award. He was tenured at Purdue University in 1997 and moved to the University of Florida in 1998 where he became full professor.

Dr. Peters serves as associate editor for the journal CAGD, APNUM, ACM ToG, GMOD and on various program committees as well as chair of the SIAM interest group on geometric design. His students have built useful tools as BezierView and TIPS.

Thesis presentations

Time: 22.05.2012, 16:30-17:00
Location: Room 072 in the Fraunhofer IGD, Fraunhoferstrasse 5, S3|05
Speaker: Michael Schmitt (Advisor Dominik Wodniok)
Title: "Estimation of Matching Probabilities for Histogrammed Intensity Patches" (Master thesis)
Abstract: (Talk will be given in German) This work presents a way to estimate the matching probabilities for histogrammed intensity patches by approximating the robustness and distinctiveness of a feature with beta distributions. We propose a simplified analytical approach to calculate the distinctiveness and show that it produces at least as good results as a specifically trained neural network. The matching probability correlates stronger with the inlier ratio than the dissimilarity score does. We then evaluate a basic tracking by detection pipeline where no longer the closest but the most probable match is chosen. Our tests show, that this improves the average performance of the system in terms of tracking rate.
Diese Arbeit stellt einen Weg vor, um die Matching-Wahrscheinlichkeiten von Histogrammed-Intensity- Patches zu berechnen in dem die Robustheit und Unterscheidungskraft der Merkmale mit Betaverteilungen angenähert werden. Wir schlagen weiterhin einen vereinfachten analytischen Ansatz zur Berechnung der Unterscheidungskraft vor und zeigen, dass dieses zumindest so guten Ergebnisses wie ein speziell für die Vorhersage der ursprünglichen Unterscheidungskraft trainiertes Neuronales Netzwerk erzielt. Die Matching Wahrscheinlichkeit korreliert stärker mit dem Anteil an Ausreißern als es die Dissimilarity-Score tut. Schließlich evaluieren wir ein Tracking-by-Detection-Verfahren, bei dem nicht mehr die nächste Übereinstimmung sondern die wahrscheinlichste gewählt wird. Unsere Testergebnisse zeigen, dass das durchschnittliche Abschneiden dieses Systems in Bezug auf Tracking-Rate verbessert wurde.

Time: 21.05.2012, 16:00-16:30
Location: Room 048 in the Fraunhofer IGD, Fraunhoferstrasse 5, S3|05
Speaker: Konstantin Fuchs (Advisor Meike Becker)
Title: "Segmentierung des Gesichtsnervs mit Active Appearance Modellen" (Master thesis)
Abstract: A large-scale invasive approach is today’s standard technique for temporal bone surgery, although it is time-consuming and causes significant tissue damage. One could overcome these disadvantages with a minimally-invasive operation method, where the whole surgery is performed through three drill canals. In order not to damage vital structures by drilling the canals, a detailed and exact operation planning is crucial, including the segmentation of all critical structures in preoperative computer tomographic data. The facial nerve is one of the important collision structures in the temporal bone region. Its segmentation is a challenging task because of its small size, its weak contrast to adjacent structures and large inter-patient variations. In this thesis, a semi-automatic two-step algorithm for the segmentation of the facial nerve is presented. We propose an Active Appearance Model based method for the extraction of the facial nerve’s centerline and evaluate four different texture descriptors in this context. For the subsequent full structure segmentation, we introduce a ray-based approach that uses the centerline for initialization. The approaches for both centerline extraction and full structure segmentation yield reliable results. They show the best segmentation quality using an introduced intensity histogram as texture descriptor.

GRIS Kolloquium

Time: 30.04.2012, 12:00
Location: Room 074 in the Fraunhofer IGD, Fraunhoferstrasse 5, S3|05
Speaker: Prof. Paul Debevec
USC Institute for Creative Technologies
Title: "From Spider-Man to Avatar, Emily and Benjamin: Achieving Photoreal Digital Actors"
Abstract: Somewhere between "Final Fantasy" in 2001 and "The Curious Case of Benjamin Button" in 2008, digital actors crossed the "Uncanny Valley" from looking strangely synthetic to believably real. This talk describes some of the technological advances that have enabled this achievement. For an in-depth example, the talk describes how high-resolution face scanning, advanced character rigging, and performance-driven facial animation were combined to create "Digital Emily", a collaboration between the USC ICT Graphics Laboratory and Image Metrics. Actress Emily O'Brien was scanned in Light Stage 5 in 33 facial poses at the resolution of skin pores and fine wrinkles. These scans were assembled into a rigged face model driven by Image Metrics' video-based animation software, and the resulting photoreal facial animation premiered at SIGGRAPH 2008. Additional examples from the motion pictures Benjamin Button, Avatar, and TRON: Legacy will also be included. The talk will also present techniques which can allow digital characters to leap from the movie screen and into the space around us, including a 3D teleconferencing system that uses live facial scanning and an autostereoscopic display to transmit a person's face in 3D and make eye contact with remote collaborators.

Bio: Academy Award winner Paul Debevec is a research professor at the University of Southern California and the associate director of graphics research at USC’s Institute for Creative Technologies. Debevec’s Ph.D. thesis (UC Berkeley, 1996) presented Façade, an image-based modeling and rendering system for creating photoreal architectural models from photographs. Using Façade he led the creation of virtual cinematography of the Berkeley campus for his 1997 film "The Campanile Movie" whose techniques were used to create virtual backgrounds in "The Matrix". Subsequently, Debevec pioneered high dynamic range image-based lighting techniques in his films "Rendering with Natural Light" (1998), "Fiat Lux" (1999), and "The Parthenon" (2004); he also leads the design of HDR Shop, the first high dynamic range image editing program. At USC ICT, Debevec has led the development of a series of Light Stage devices for capturing and simulating how objects and people reflect light. They have been used to create photoreal digital actors in films such as "Spider-Man 2", "Superman Returns", "The Curious Case of Benjamin Button", and "Avatar". He received ACM SIGGRAPH’s first Significant New Researcher Award in 2001 and co-authored the 2005 book "High Dynamic Range Imaging". In addition, he chaired the SIGGRAPH 2007 Computer Animation Festival and is Vice President of ACM SIGGRAPH. His 2010 Scientific and Engineering Academy Award® was presented to him, Tim Hawkins, John Monos, and Mark Sagar for their work on the Light Stage facial capture systems.

Thesis presentations

Time: 24.04.2012, 17:15-18:00 Uhr
Location: Room 074 in the Fraunhofer IGD, Fraunhoferstrasse 5, S3|05
Speaker: Leander Baumann (Betreuer Dominik Wodniok)
Title: "Using Human Visual Intelligence - Multiple Sequence Alignment as Serious Game" (Bachelor thesis)
Abstract: (Talk will presumably be given in German) In der Bioinformatik werden häufig sog. multiple Sequenzalignments benötigt, deren Erstellung meist sehr rechenintensiv ist. Die Erfahrung zeigt, dass ihre Qualität oft nicht so gut wie erwünscht ist und dass interessanterweise Menschen mittels ihrer Mustererkennungsfähigkeiten fähig sind, in computererstellten Alignments sowohl lokale als auch globale Optimierungen vorzunehmen.

In dieser Arbeit soll diese menschliche Fähigkeit genutzt werden, indem die Aufgabe des Sequenzalignens Menschen als Spiel präsentiert wird. Die Form des Spiels soll dabei den Spass am Wettbewerb ansprechen. Vergleichbare Systeme nutzen bereits mit unterschiedlichem Erfolg menschliche Intelligenz, um Probleme zu lösen, die für Computer als schwierig gelten.

Inhalt dieser Bachelorarbeit ist die Erstellung eines Prototypen eines auf einem Webframework basierenden Serious Games zur Verbesserung computererstellter Sequenzalignments, sowie eine einfache, initiale Evaluierung des Systems und der erste Vergleich der menschlichen Leistungsfähigkeit beim Erstellen von Alignments im Vergleich mit rein von Computern erstellten Alignments.

GRIS Kolloquium

Time: 11.04.2012, 14:00
Location: Room 074 in the Fraunhofer IGD, Fraunhoferstrasse 5, S3|05
Speaker: Rafael Saracchini, Ph.D.
Laboratory of Vision Informatics, State University of Campinas
Title: "Robust Methods for Example-Based Photometric Stereo"
Abstract: In several areas of knowledge, such as in archeology, industrial quality control, medicine and security, there is a strong need for analysis of the 3-dimensional shape of objects. Most 3D scanners are bulky, expensive and not always provide a suitable solution or offer the required precision. Photometric Stereo is a promising 3D capture technology that does not have such limitations, as it can recover surface normals and albedo using only light sources and cameras. In this talk will be presented methods for Example-Based Photometric Stereo and gradient integration aiming to be sufficiently reliable for use in practical applications and using low-cost devices such as consumer cameras, mobile phones and portable computers.

Thesis presentations

Time: 27.02.2012, 16:45-17:15
Location: Room 072 in the Fraunhofer IGD, Fraunhoferstrasse 5, S3|05
Speaker: Tobias Große-Puppendahl (Advisor Andreas Braun)
Title: "Multi-hand interaction using custom capacitive proximity sensors" (Master thesis)
Abstract: The recognition of gestures in free space using sensors that determine the proximity of a body mass based on electric field variance is a challenging research topic. Arrays of such capacitive proximity sen-sors allow the creation of novel user interaction systems based on recognizing presence and position of body parts and inferring performed gestures in three dimensions. These systems may be incorporated as unobtrusive remote control in home automation scenarios or automotive applications, for example as a smart car-dashboard. Present systems that use time-division multiplexing have limitations in their temporal and spatial resolution. In this thesis a novel capacitive proximity sensor system that uses a combination of frequency division and time division multiplex is presented, improving both temporal and spatial resolution. With this affordable system one is able to detect fast multi-hand gestures in three dimensions above large surface areas. Moreover, a method for object recognition using capacitive proximity sensors is extended and refined and a new object tracking method, employing particle filters is presented. A user evaluation emphasizes the feasibility of the presented capacitive sensing system as an explicit interaction modality.

Time: 27.02.2012, 16:00-16:30
Location: Room 072 in the Fraunhofer IGD, Fraunhoferstrasse 5, S3|05
Speaker: Mohammad H. Keyhani (Advisor Stefan Wesarg)
Title: "Super-Resolution from a Single Medical 3D Image Data Set" (Master thesis)
Abstract: Super-Resolution Reconstruction is a technique for recovering and reconstructing a higher resolution of a low-resolution image. There exist classical and example-based super-resolution reconstruction methods, and especially a combination of them leads to reasonable results. In this thesis, we combine a classical and an example-based super-resolution reconstruction method for enhancing the resolution of medical image data sets. Moreover, we present a super-resolution reconstruction method for solving the anisotropy problem in CT and MRI 3D image data sets. Our patch-based method improves the quality of low-resolution structures in anisotropic slices using high-quality structures present in isotropic slices of the data set. Consequently, an isotropic super-resolution image is calculated for each anisotropic slice of the data set. For our approach, no further images with subpixel misalignments are necessary as usually needed by classical superresolution strategies, but the information available in the 3D data set suce for calculating reasonable super-resolution images. Various experiments with real medical data sets prove the eciency and correctness of our implementation. We are successful in approximating the super-resolution of medical images reasonably, and improving their quality signicantly.

Time: 22.02.2012, 12:00-12:30
Location: Room 242 in the Fraunhofer IGD, Fraunhoferstrasse 5, S3|05
Speaker: Nazli Bozoglu (Advisor Sebastian Steger)
Title: "Evaluierung von Algorithmen zur Segmentierung von Lymphknoten in CT Datensätzen" (Master thesis)
Abstract: (Talk will be held in German) Die Untersuchung der Größe und Form von Lymphknoten spielt in der Diagnostik und Verlaufskontrolle von Krebspatienten eine wichtige Rolle. Für diese Untersuchung müssen die Lymphknoten in Computertomographie (CT) Datensätzen segmentiert werden. Die manuelle Segmentierung ist zeitaufwändig und fehleranfällig. Für eine präzise Segmentierung von Lymphknoten sind automatische Segmentierungsverfahren erwünscht. In der Literatur werden einige vielversprechende Segmentierungsalgorithmen beschrieben. Alle diese Algorithmen haben ihre Vor- und Nachteile. Eine Gegenüberstellung der Algorithmen wird vermisst.

In dieser Arbeit werden vier neue automatische Segmentierungsalgorithmen entwickelt. Die Parameter der Verfahren werden optimiert. Anhand Goldstandards von Experten werden im Anschluss diese vier Methoden und regionen-, modell- und strahlenbasierte
Segmentierungsalgorithmen aus der Literatur in Bezug auf die Genauigkeit, Laufzeit und Robustheit evaluiert und gegenübergestellt. Für die Evaluierung steht ein CT-Datensatz zur Verfügung, für den manuelle Segmentierung von Experten vorliegen. Der Datensatz besteht aus 49 Lymphknoten der Kopf und Halsregion von Patienten mit Mundhöhlenkarzinomen. Bei der Gegenüberstellung der Algorithmen wurde festgestellt, dass der strahlenbasierteAnsatz, bei dem die Segmentierung unter Berücksichtigung lokaler Gradienteneigenschaften und Intensitätseigenschaften von Lymphknoten mit anschließender globaler Optimierung erfolgt, am besten für die Segmentierung von Lymphknoten geeignet ist. Die Laufzeit der Segmentierung ist mit 8 Sekunden pro Lymphknoten 3-5 fache der anderen Segmentierungsverfahren. Die Segmentierung ist aber sehr präzise. Der durchschnittliche DSC beträgt 0.80. Bei der Initialisierung mit unterschiedlichen Saatpunkten ist die Methode sehr robust. Bei der Segmentierung von vergrößerten Lymphknoten beträgt der durchschnittliche DSC 0.86, bei der Segmentierung von nekrotischen Lymphknoten 0.81.

Anhand von automatischen Segmentierungsalgorithmen können Lymphknoten erfolgreich segmentiert werden. Die Ergebnisse der Segmentierug können z.B. für die dreidimensionale Visualisierung, Verlaufskontrolle von Krebspatienten oder Operationsplanung weiterverarbeitet werden.

GRIS Kolloquium

Time: 08.02.2012, 16:30
Location: Room 073 in the Fraunhofer IGD, Fraunhoferstrasse 5, S3|05
Speaker: Ryusuke Sagawa, Ph.D.
Intelligent Systems Research Institute, AIST
Title: "One-shot Dense 3D Reconstruction for Objects in Fast Motion"
Abstract: 3D scanning of moving objects has many applications, for example, marker-less motion capture, analysis on fluid dynamics, object explosion and so on. One of the approaches to acquire accurate shape is a projector-camera system, especially the methods that reconstruct a shape by using a single image with a static pattern are suitable for capturing fast moving objects.

In this research, we propose a method that uses a grid pattern consisting of sets of parallel lines. The pattern is spatially encoded by a periodic color pattern. While information is sparse in the camera image, the proposed method extracts the dense (pixel-wise) phase information from the sparse pattern. As the result, continuous regions in the camera images can be extracted by analyzing the phase. Since there remain one DOF for each region, we propose the linear solution to eliminate the DOF by using geometric information of the devices, i.e. epipolar constraint. In addition, the solution space is finite because the projected pattern consists of parallel lines with equal intervals, and the linear equation can be efficiently solved by integer least square methods. In the experiments, a scanning system that can capture an object in fast motion has been actually developed by using a high-speed camera. In the experiments, we show the sequence of dense shapes of an exploding balloon, and other objects at more than 1000 fps.

Bio: Ryusuke Sagawa is a researcher at Service Robotics Research Group, Intelligent Systems Research Institute, National Institute of Advanced Industrial Science and Technology (AIST), Japan. He received a BE in Information Science from Kyoto University in 1998. He received a ME in Information Engineering in 2000 and Ph.D. in Information and Communication Engineering from the University of Tokyo in 2003. He was an assistant professor at the Institute of Scientific and Industrial Research, Osaka University. He stayed at ETHZ as a visiting researcher in 2008 and moved to AIST in 2010. His primary research interests are computer vision, computer graphics and robotics (mainly geometrical modeling and visualization).

Thesis presentations

Time: 18.01.2012, 12:00-12:30
Location: Room 048 in the Fraunhofer IGD, Fraunhoferstrasse 5, S3|05
Speaker: Simon Diel (Advisor Tatiana von Landesberger)
Title: "Visual Analysis of Default Effects in Networks" (Bachelor thesis)
Abstract: The default of Lehman Brothers lead to a wide-spread financial crisis. This is caused by the strong interconection of financial institutions around the world. In global financial networks, a default of one company leads to a chain reaction of other companies defaulting. These effects need to be analyzed in order to be able to assess stability of financial systems and systemic relevance of individual financial entities. The goal of the thesis was to develop new Visual Analytics tools that support this assessment - the simulation of default effects and their comparison in various situations. In this way, stability of financial systems can be analyzed and predictions of default effects can be performed.

GRIS Kolloquium

Time: 09.12.2011, 14:00
Location: Room 073 in the Fraunhofer IGD, Fraunhoferstrasse 5, S3|05
Speaker: Prof. Paul Fieguth
University of Waterloo
Title: "Hierarchical Discrete-State Random Fields"
Abstract: There exist many applications for large-scale, discrete-state Markov or Gibbs random fields. In particular, our research has been driven by the demand for methods to synthesize large images of porous media (cement, limestone, bone, cartilage etc.). Because such media contain structures over a wide range of length scales, there are significant difficulties in generating samples using statistical techniques such as simulated annealing.

Hierarchical methods have previously been found quite effective for such problems. In this talk I will present a hierarchical approach to discrete state fields that offers over an order of magnitude reduction in computational complexity versus existing hierarchical techniques.

Finally, given an efficient prior model it should be possible to use it in other statistical tasks, such as data fusion or reconstruction.

Bio: Paul W. Fieguth received the B.A.Sc. degree from the University of Waterloo, Ontario, Canada, in 1991 and the Ph.D. degree from the Massachusetts Institute of Technology, Cambridge, in 1995, both degrees in electrical engineering.

He joined the faculty at the University of Waterloo in 1996, where he is currently Professor and, since 2010, Department Chair in Systems Design Engineering. He has held visiting appointments at the University of Heidelberg in Germany, at INRIA/Sophia in France, at the Cambridge Research Laboratory in Boston, at Oxford University and the Rutherford Appleton Laboratory in England, and with postdoctoral positions in Computer Science at the University of Toronto and in Information and Decision Systems at MIT. His research interests include statistical signal and image processing, hierarchical algorithms, data fusion, and the interdisciplinary applications of such methods, particularly to remote sensing. He is the author of "Statistical Image Processing and Multidimensional Modeling" from Springer.

Thesis presentations

Time: 06.12.2011, 14:45-15:30
Location: Room 048 in the Fraunhofer IGD, Fraunhoferstrasse 5, S3|05
Speaker: Ralf Gutbell (Advisors Wissam El Hakimi, Meike Becker)
Title: "Development of a surgical planning software for drilling at the lateral skull base" (Master thesis)
Abstract: Research in the field of minimally-invasive surgery is important, because it will allow surgeries at lower risk levels and will help to safe money and time as well. Several critical structures are located at the lateral skull base, so the minimally-invasive approach has not been performed in this area yet. Small errors in the drill placement could lead to severe complications for the patient. In this thesis a prototype for multi-port planning is developed, which models the critical structures and considers the placement inaccuracy of the drill. It also allows the surgeon to determine the drill’s size and a minimal distance to the critical structures, which the drill paths have to respect. A prototype has been developed, which employs and extends the medical simulation framework SOFA. New models have been created, describing the drill paths and collision components, enabling the collision detection and its visualization. Another feature of the prototype is the option of simulating the drill feed for a single drill path. The simulation models the drill and its movement from the entry point to the surgical target. This option allows the user assess the spatial relation between the instrument and the critical structures. The prototype’s visualization features, its results and performance are analyzed for different drill settings, datasets and computer systems. The drill settings are combinations of drill sizes, minimal distance and drill placement inaccuracies. The evaluation shows that the created collision model has been successfully integrated in SOFA’s collision pipeline. The implemented visualization features adequately represents the collisions and drill path parameters. The prototype’s performance is limited and needs to be improved. Different solutions to achieve that and possible new features are presented.

Time: 06.12.2011, 14:00-14:45
Location: Room 048 in the Fraunhofer IGD, Fraunhoferstrasse 5, S3|05
Speaker: Steffen Herbert (Advisors Wissam El Hakimi, Meike Becker)
Title: "Super-Resolution Ansatz zur medizinischen Bildverbesserung mittels Unsicherheitsmodellen" (Master thesis)
Abstract: The resolution and quality of magnetic resonance imaging (MRI) and computer tomography (CT) plays an important role in the diagnosis of e.g. heart anomalies and brain tumors. Images with a higher resolution lead thereby to significantly less wrong decisions. However, MR and also CT images often suffer from a highly anisotropic resolution. We propose a super-resolution reconstruction (SRR) approach, based on a 3D space-variant interpolation method, for reconstructing a high resolution (HR) image from different orthogonal low resolution (LR) images. Thereby, uncertainties of voxels, which arise during image acquisition and preprocessing, are considered. We formulate a space-variant interpolator in an efficient way and adapt the algorithm to the structure of the interpolation regions in order to allow an efficient parallel processing. Furthermore, we propose an iterative scheme of the SRR algorithm where small regions are independently processed until they are reconstructed. Experiments with synthetic data and real data reveal high contrast and sharp details of the SR result, especially in regions of small object structures. Compared to the average of the upsampled LR images, we obtain a significant improvement in terms of image quality and contrast. We provide an extensive evaluation of the SRR process in order to optimize the algorithm and to estimate its parameters in an adaptive way. Furthermore, the experiments show that the algorithm is able to handle minor intensity differences of the input images, which arise in real scans.

Time: 28.11.2011, 13:45-14:15
Location: Room 048 in the Fraunhofer IGD, Fraunhoferstrasse 5, S3|05
Speaker: Andreas Lassmann (Advisor Matthias Kirschner)
Title: "Simulation der Festigkeit von Pedikelschrauben in Wirbel" (Bachelor thesis)
Abstract: (Talk will be held in German) Diese Bachelorarbeit behandelt die Simulation der Festigkeit von Pedikelschrauben in Wirbeln. Ich gebe einen Einblick in das medizinische Umfeld und leite daraus ein physikalisches Modell ab. Damit soll geklärt werden, ob sich die Haltekraft der Pedikelschraube anhand der Knochendichte bestimmen lässt. Zur Lösung des Problems wird ein Simulationswerkzeug verwendet, mit dessen Hilfe eine Finite Elemente Analyse durchgeführt werden kann. Die dazu benötigte Finite Elemente Methode wird für die Berechnung der Kraft modifiziert. Zusätzlich wird eine Strategie zum Stoppen der Simulation umgesetzt. Ich zeige mein Vorgehen während der Simulation auf und diskutiere deren Ergebnisse. Bei der Evaluation werden sich Schwankungen und Ausreißer abzeichnen, die im Weiteren diskutiert werden. Zuletzt zeige ich eine alternative Simulationsmethode auf und gebe einen Ausbilck auf Verbesserungsmöglichkeiten.

Time: 28.11.2011, 13:00-13:30
Location: Room 048 in the Fraunhofer IGD, Fraunhoferstrasse 5, S3|05
Speaker: Patrick Hörmann (Advisor Matthias Kirschner)
Title: "Erstellen Statistischer Formmodelle mit Hilfe des EM-ICP-Algorithmus" (Bachelor thesis)
Abstract: (Talk will be held in German) In der aktuellen medizinischen Bildverarbeitung werden bildgebende Verfahren verwendet, um das Innere des Menschen sichtbar zu machen. Da aneinanderliegende Organe eine ähnliche Beschaffenheit haben und deshalb einen geringen Kontrast zueinander aufweisen, ist es schwierig, diese rein intensitätsbasiert zu segmentieren. Zur robusten Segmentierung werden deshalb statistische Formmodelle (SFM) eingesetzt, die Vorwissen über die Verformung eines Organes in den Segmentierungsprozess einfließen lassen. Häufig repräsentiert man solche Datensätze durch Punktwolken, deren Punkte zueinander in Korrespondenz gesetzt werden. Um nun diese Korrespondenzen aufzustellen, wird ein Registrierungsalgorithmus benötigt, wie zum Beispiel der EM-ICP Algorithmus. In dieser Bachelorarbeit werde ich auf das Erstellen statistischer Formmodelle mit Hilfe des Expectation Maximisation Iterative Closes Point Algorithmus eingehen. Die so erhaltenen SFMs werde ich dann mit SFMs vergleichen, die mit diversen anderen Methoden erstellt wurden.

Time: 28.11.2011, 10:00-10:30
Location: Room 073 in the Fraunhofer IGD, Fraunhoferstrasse 5, S3|05
Speaker: Adwait Gandhe (Advisor Martin Ritz)
Title: "Dense 3D Reconstruction and Object Recognition Using a Minimum Set of Inside-Out Images" (Bachelor thesis)

Time: 25.11.2011, 10:45-11:15
Location: Room 324 in the Fraunhofer IGD, Fraunhoferstrasse 5, S3|05
Speaker: Galin Bobev (Advisor Stefan Roth)
Title: "Schätzung von Kartendarstellungen für automotive Anwendungen durch 2D Scene Labeling" (Bachelor thesis)

Time: 25.11.2011, 10:00-10:30
Location: Room 324 in the Fraunhofer IGD, Fraunhoferstrasse 5, S3|05
Speaker: Christoph Dann (Advisor Stefan Roth)
Title: "A Spatial Consistent CRF for Semantic Image Segmentation" (Bachelorthesis)

GRIS Colloquium

Time: 07.11.2011, 10:00
Location: Room 073 in the Fraunhofer IGD Fraunhoferstrasse 5, S3|05
Speaker: Cornelia Denk
Realtime Technology AG
Title: "At the verge of change - How High-end Visualization drives industrial decision-making"
Abstract: Beyond the film and gaming industry, High Performance Graphics has found it‘s way into industrial decision-making processes on a broad scale – from the first design up to the point of sale.

These industry applications, however, come with a different set of challenges:
-Where to draw the line between high performance visualization and simulation?
-How to digitize and render real, physical samples efficiently?
-How to find the balance between required process optimization and freedom of creativity? -How to combine different specialized algorithms to meet divergent requirements?
-Which new data models and asset standards have to be developed as a result?

I will give you insight on how to solve these challenges and share our vision about opportunities to take high performance visualization to the next level of enterprise applications.

Beginning in 2005, Cornelia Denk worked as a researcher on pre-computed global illumination algorithms for RTT. From 2008 to 2010 she continued her work as Research Manager and was in charge of defining the technical visualization strategy. As Manager Technology & Innovation she is responsible for building up strategic partnerships with industry accounts and universities. Her role also involves technical trends and innovation strategy.

Time: 11.10.2011, 16:30
Location: Room 074 in the Fraunhofer IGD Fraunhoferstrasse 5, S3|05
Speaker: Jun.-Prof. Dr. Martin Fuchs
Visualisierungsinstitut der Universitaet Stuttgart
Title: "The Visual Computing Cycle"
Abstract: Visual computing encompasses a multitude of subjects which have evolved around a common goal: creating and understanding worlds with pictures. In image space, the computer graphics and vision pipelines form the visual computing cycle: linking the acquisition and the rendering of real-world scenes facilitates novel interactions between human beings, computers, and the real world.

Recent years have brought forth the emergence of computational photography, which surpasses the paradigms of traditional cameras after decades of virtually unchanged designs, and appearance synthesis which fabricates materials with looks that can not just be designed, but actually be programmed through algorithmically guided manufacturing.

In this talk, I will present a selection of works on the visual computing cycle that I contributed to in the last years and discuss the outlook for my newly founded working group on Visual Computing at VISUS, the visualization research center of Stuttgart University.

Bio: Martin Fuchs did his PhD at the MPI/Saarland University and received his degree in 2008. Afterwards, he was a Postdoc at Princeton University from 09/2009 till 01/2011. Currently, he holds a position as junior professor at the Visualisation Research Center at the University of Stuttgart.

Time: 26.09.2011, 16:00
Location: Room 074 in the Fraunhofer IGD Fraunhoferstrasse 5, S3|05
Speaker: Carsten Rother, Ph.D.
Computer Vision Group (Microsoft Research Cambridge, UK)
Title: "Recent Progress in Stereo Matching"
Abstract: Stereo matching is the task of reconstructing a 3D scene from a pair of stereo images. This is a long-standing goal in computer vision and a large number of approaches exist. In this presentation I will review various works we have done in this field which were presented at recent conferences, CVPR '11 and BMVC '11. One line of research is to develop a rather complex objective function which jointly models depth and the collection of 3D objects that constitute the scene. The other line of research investigates fast techniques which still achieve excellent results in practice. In particular I will present the stereo guided filter and stereo PatchMatch algorithm.
Bio: Carsten Rother received the Diploma degree with distinction in 1999 from the University of Karlsruhe. He did his PhD at the Royal Institute of Technology Stockholm, supervised by Stefan Carlsson and Jan-Olof Eklundh. His thesis was nominated for the Best Nordic Thesis Award 2003-2004, as one out of two candidates from Sweden. From 2003-04, he was PostDoc at Microsoft Research Cambridge, and since 2004 has been a permanent researcher there. His research interests are in the field of “(Markov) Random Field Models for Computer Vision”, “Discrete Optimization”, and “Vision for Graphics” (in particular interactive segmentation and matting). He has published more than 20 high impact papers (at least 10 citations) at international conferences and journals, and co-authored a paper which won the best paper honourable mention award at CVPR’05. He serves on the program committee of major conferences (e.g. SIGGRAPH, ICCV, ECCV, CVPR, NIPS), and has been area chair for BMVC‘08, and ‘09. He has organized some workshops (“Interactive computer vision” at ICCV’07, “Color and Reflectance” at ICCV’09), and supervises several PhD students.

Thesis presentations

Time: 30.05.2011, 13:30-14:00
Location: Room 072 in the Fraunhofer IGD, Fraunhoferstrasse 5, S3|05
Speaker: Patrick Muecke (Advisor Domik Wodniok)
Title: "3D Surface Reconstruction from Multi-Resolution Depth Maps" (Diploma thesis)
Abstract: In this thesis, a new method for surface reconstruction from depth maps is being proposed and analyzed. Due to the acquisition method, surface samples of depth maps normally are no ideal point samples, but rather represent an area of the represented surface. Still, this footprint information inherent in depth map samples has not been explicitly utilized by any of the many different surface reconstruction methods by now. The proposed method tries to improve the reconstruction quality by  explicitly  taking  the  sample  footprint  information  into  account. Furthermore,  the  proposed  method  is  intended  to  work  on  depth maps generated by Multi-View Stereo and was hence designed to be robust against noise and incompleteness of sampling. Multi-resolution reconstruction was implemented in order to be able reconstruct fine details in large scenes.

Time: 09.05.2011, 14:00-14:30
Location: Room 073 in the Fraunhofer IGD, Fraunhoferstrasse 5, S3|05
Speaker: Lilli Gong (Advisor Stefan Wesarg)
Title: "Analysis of the mitral valves based on high resolution MR images" (Bachelor thesis)

Time: 12.04.2011, 15:00-15:30
Location: Room 072 in the Fraunhofer IGD, Fraunhoferstrasse 5, S3|05
Speaker: Bastian Moldenhauer (Advisor Daniel Weber)
Title: "Collision Detection for Deformable Bodies" (Diploma thesis)

Time: 28.03.2011, 17:00-17:30
Location: Room 074 in the Fraunhofer IGD, Fraunhoferstrasse 5, S3|05
Speaker: Jeanette Forster (Advisor Kawa Nazemi)
Title: "Semantic visualization of searchresults based on Linked Open Data" (Bachelor thesis)
Abstract: (Talk will be given in German) Semantically annotated data gain more and more importance in future information acquiring processes. Especially the Linked Open Data (LOD) format has experienced a great growth. However, the user-interfaces of most webapplications still cannot re_ect the added value of semantics data. The following work describes a new approach of user-centered data-adaptive semantics visualization for searchresults based on Linked Open Data applications. The developed concepts make use of the advantages of semantics data combined with an adaptive composition of information visualization techniques. The work starts with a description of basic concepts for information visualization and semantics data. A discussion of advantages and disadvantages of visualization techniques especially for semantics data follows. The following chapter describes existing visualization techniques and important LOD systems. An analysis and comparison of applications especially for visualization of semantics data reveal existing problems with current approaches. After that, the new approach will bring the Linked Open Data and the visualization characteristics together and introduces a visualization system that adapts the composition of visualizations based on the underlying data structure. The next section describes the implementation of a system prototyp, which shows the advantages of the new strategies. A summery of the developed concepts and a perspective for future research themes conclude this work.

Time: 28.03.2011, 16:30-17:00
Location: Room 074 in the Fraunhofer IGD, Fraunhoferstrasse 5, S3|05
Speaker: Tsvetoslava Vateva (Advisor Sami ur Rahman)
Title: "Evaluating a super resolution reconstruction algorithm and enhancing image quality using simulated annealing technique" (Bachelor thesis)

GRIS Colloquium

Time: 01.02.2011, 16:30
Location: Room 074 in the Fraunhofer IGD Fraunhoferstrasse 5, S3|05
Speaker: Gianpaolo Palma
University of Pisa
Title: "Surface appearance reconstruction from video sequences"
Abstract: (Talk will be held in English) The technologies for the 3D digitalization of Cultural Heritage artworks have been improved in a considerable manner in the last few years. While the methodologies for the acquisition of the shape through 3D scanning allow the reconstruction of very accurate geometries in an affordable times, on the other hand the acquisition of the object’s surface appearance is more complex and the technology is in a more primitive status. We propose a new technique for the acquisition of the surface appearance of an object from video sequence and its projection over a 3D scanning triangular mesh of the same object. The technique is composed by two steps: the registration of the video sequence over the mesh by calibration of the camera; the reconstruction of the surface appearance with different degree of quality.

The goal of the first step is to obtain an accurate alignment of each frame of the video over the mesh to allow the bidirectional data transfer between them. Our solution uses two different approaches: feature-based registration by KLT video tracking, and statistic-based registration by maximizing the Mutual Information between the gradient map of the frame and the gradient map of the rendering of the 3D model with some illumination related properties, such as surface normals and ambient occlusion.

The goal of the second step is to reconstruct the reflectance behavior of the surface by statistical analysis of the color samples projected in each point from each frame of the video.

Thesis presentations

Time: 24.01.2011, 14:00-14:30
Location: Room 072 in the Fraunhofer IGD, Fraunhoferstrasse 5, S3|05
Speaker: Martin Hess (Advisor Sebastian Bremm)
Title: "Visueller Vergleich hierarchisch organisierter Daten" (Bachelor thesis)

GRIS Colloquium

Time: 09.12.2010, 16:30
Location: Room 074 in the Fraunhofer IGD Fraunhoferstrasse 5, S3|05
Speaker: Prof. Klara Nahrstedt
University of Illinois at Urbana-Champaign
Title: "3D Tele-Immersive Environments for Everybody"
Abstract: For the last two decades, we have seen expensive, special purpose, and complex tele-immersive multimedia environments over high-performance infrastructures such as CAVE, Grid computing, and high-speed leased networks that only few could afford. With the advances in COTS 3D cameras, 3D displays, multi-core, multi-GPU computing, large increases in ubiquitous network bandwidth availability over Internet (wired and wireless), 3D tele-immersive environments for everybody are becoming possible. However, with the availability of COTS 3D resources, new challenges arise such as how do we organize efficiently resources within each tele-immersive room, how do we disseminate and deliver multiple correlated sources per room efficiently across multiple participating rooms using current COTS computing and networking infrastructures, and how do we give users the power of control to manipulate the shared virtual 3D environments.

In this talk, I will present the TEEVE framework which explores multiple 3D cameras and displays resources in tele-immersive rooms, includes a novel cross-layer QoS-adaptive multicasting architecture with view-based stream dissemination, called View-Casting, and gives user Wii-based view control of 3D video and graphics objects. The TEEVE framework aims for (1) effective and adaptive coordination of resources, synchronization and soft QoS-enabled delivery of tele-immersive visual streams to remote rooms, (2) effective view-casting model for different view dissemination in the multi-party 3D tele-immersive environments, and (3) easy view control by users. I will discuss various algorithmic, protocol, architectural solutions and user-studies feedback of the TEEVE framework that presents an instance of the next generation of ubiquitous tele-immersive environments for dancers, gaming users and others. The current set of TEEVE experimental results confirms that we are moving in the right direction in terms of our design and approaches.

Bio: Klara Nahrstedt is a professor at the University of Illinois at Urbana-Champaign, Computer Science Department. Her research interests are directed towards tele-immersive systems and applications, Quality of Service (QoS) management in wired and wireless networks, Quality of Experience, QoS routing, QoS-aware resource management in distributed multimedia systems, and pervasive mobile multimedia systems and applications. She is the coauthor of the widely used multimedia books "Multimedia: Computing, Communications and Applications" published by Prentice Hall, and "Multimedia Systems" published by Springer Verlag. She is the recipient of the Early NSF Career Award, the Junior Xerox Award, the IEEE Communication Society Leonard Abraham Award for Research Achievements, the University Scholar Award, the Humboldt Research Award, and the Ralph and Catherine Fisher Professorship Chair. She was the editor-in-chief of ACM/Springer Multimedia Systems Journal (2000-2005), she was the general co-chair of ACM Multimedia 2006, the general chair of ACM NOSSDAV 2007, the general chair of IEEE Pervasive Computing and Communications (Percom) 2009, and she is currently the elected chair of ACM Special Interest Group in Multimedia (2007-2011). Klara Nahrstedt received her BA in mathematics from Humboldt University, Berlin, in 1984, and M.Sc. degree in numerical analysis from the same university in 1985. She was a research scientist in the Institute for Informatik in Berlin until 1989. In 1995 she received her PhD from the University of Pennsylvania in the Department of Computer and Information Science. She is the member of ACM and IEEE Fellow.

Inaugural lecture

Time: 07.12.2010, 14:00
Location: Room 074 in the Fraunhofer IGD Fraunhoferstrasse 5, S3|05
Speaker: Priv.-Doz. Dr. Arjan Kuijper
Fraunhofer IGD
TU Darmstadt
Title: "Scales, images, and shapes"
Abstract: Humans are extremely good in analyzing the contents of images and interpreting shapes – even without a priori knowing them. In computer vision probably the most challenging task is therefore to come up with mathematical models that mimic the human visual system. An essential requirement is to understand the models themselves. This requires knowledge of various mathematical fields, like topology, geometry, partial differential equations, and singularity theory. In my talk a basic element for proposed methods is discussed: scale. This leads to the mathematical concepts of scale space for images and symmetry sets for shapes. I will briefly introduce these concepts and discuss their ability to represent the underlying structure, as these models capture essential characteristics of images and shapes. In the last part of the talk I will sketch connections of this research area "Image Processing and Analysis" with the existing Visual Computing environment in Darmstadt.

GRIS Colloquium

Time: 30.11.2010, 16:30
Location: Room 074 in the Fraunhofer IGD Fraunhoferstrasse 5, S3|05
Speaker: Dr. Arno Zinke
University of Bonn
Title: "Towards Virtual Hair - Practical Physically-Based Hair Rendering"
Abstract: (Talk will be held in English) The optical simulation of human hair is a very challenging topic, which is not only relevant to CGI but currently becoming an integral part of cosmetic research. Most difficulties arise from the fact that thousands of individual hair strands have to be considered. Even worse, not only the sheer number is a problem but also a single strand is a very complex system. This presentation will be focused on recent advances in the field of physically-based hair rendering. After a gentle introduction to the topic, a review on the state-of-the-art concepts and techniques will be given followed by a brief discussion on applications in the field of CGI and cosmetics.

Bio: Arno Zinke received a PhD for the dissertation on physically-based rendering of fiber assemblies from University of Bonn in 2008. Besides his occupation as the CEO of GfaR mbH, a company who provides innovative R&D in the field of visual computing, he is also a researcher at the "Multimedia, Simulation and Virtual Reality" group (University of Bonn) and was lecturing at the University of Frankfurt and the Bonn-Aachen International Center for Information Technology. His current research focuses on modeling of fiber assemblies as well as on topics related to human motion analysis and synthesis.

Time: 23.11.2010, 16:30
Location: Room 074 in the Fraunhofer IGD Fraunhoferstrasse 5, S3|05
Speaker: Karol Myszkowski
Max Planck Institute for Informatics
Title: "Overcoming physical limitations of display devices in rendering"
Abstract: The knowledge of human visual system (HVS) enables more efficient image rendering by overcoming physical constraints of display devices. This talk presents a number of successful examples of embedding HVS models into real-time rendering pipelines. In particular, I showcase the problem of improving the appearance of highlights and light sources by boosting their apparent brightness using the temporal glare technique. Also, I discuss how to overcome physical contrast limitations of display devices by using the 3D unsharp masking technique to boost the apparent contrast. Finally, I present our technique for apparent resolution enhancement, which enables showing image details beyond the physical pixel resolution of the display device.

Thesis presentations

Time: 22.11.2010, 14:45-15:15
Location: Room 048 in the Fraunhofer IGD, Fraunhoferstrasse 5, S3|05
Speaker: Eugen Flehmann (Advisor Sami ur Rahman)
Title: "Segmentierung der Aorta, Koronararterien und Bestimmung der geometrischen Parameter" (Master thesis)
Abstract: (Talk wil be given in German) In this Master Thesis an algorithm is proposed which automatically segment the aorta and coronary arteries from the 3D MRI data and extract geometric parameters like diameter of the aorta and the coronary arteries, curvature of the aortic arch, and angle of the coronary arteries. Two sets of input images are used. One set contains the complete aorta and the second set contains coronary arteries along with some part of the aorta near coronary arteries.

In the proposed method, first the aorta and coronary arteries are registered. A straightforward registration of two images is not possible because both images share only small common area near the coronary arteries. Due to the blood circulation and breathing effect the heart chambers in both images do not remains constant. This problem is solved by roughly cutting the shared area from both images, first applying rigid registration on these sub images and then applying the registration parameters on the whole image. The rigid registration roughly aligns both images and after that, deformable registration is applied. The next Step is to segment these registered images. The seed points are selected automatically in both images and then the images are segmented using Fast Marching algorithm. The complete process is automatic. After this the aorta and coronary arteries are fused. A filter is created that takes the two segmented images as input and applies a voxel-wise mean computation. After segmentation and fusion, geometric parameters like diameter of the aorta and the coronary arteries, curvature of the aortic arch, and angle of the coronary arteries area computed. These parameters computation is important for selecting patient specific catheter selectaion during catheter angiography.

Time: 22.11.2010, 15:45-16:15
Location: Room 048 in the Fraunhofer IGD, Fraunhoferstrasse 5, S3|05
Speaker: Andreas Franek (Advisor Matthias Bein)
Title: "Genetic B-Spline Approximation with uniform and marked knots" (Bachelor thesis)
Abstract: (Talk wil be given in German) B-Splines are defined by the degree of the base functions, the knot vector and control points. To fit a suitable B-Spline to a given discrete sampled input curve is a complex optimization problem in which the above mentioned unknowns have to be determined. This also holds in the case of uniform B-Splines and marked knots which force the B-Spline to interpolate the marked control points.

In this work we are going to describe a genetic approximation method to approximately solve the optimization problem. The solution space of possible approximations is very large and can not be scanned completely. In oder to get a good result in real-time heuristics and parallelization methods are used to speed up the process. The genetic algorithm finds a good B-Spline with a calculation time of less then a second. Compared to the results of a previously used iterative algorithm, the fitted B-Spline stands out by consisting fewer control points while still having a lower approximation error.

Time: 18.11.2010, 13:00-13:30
Location: Room 073 in the Fraunhofer IGD, Fraunhoferstrasse 5, S3|05
Speaker: Fabian Langguth (Advisor Jens Ackermann)
Title: "Photometric Stereo for Outdoor Webcams" (Bachelor thesis)
Abstract: (Talk wil be given in German) Nowadays, laser scanners or other types of active capturing devices deliver good results for many kinds of objects. But especially for large-scale outdoor scenes it is often difficult or very costly to recover the geometry and materials of objects with dedicated equipment.

In this work we are going to describe a passive image based reconstruction technique designed to recover both geometry and reflectance properties of an outdoor object using a photometric stereo approach. The method is specialized for images from static outdoor webcams and therefore introduces a complete new application of the photometric stereo principle. We first calibrate the orientation of the camera and can then use the sun as direct light source in the scene. The reconstruction itself is based on a recent technique that can deal with a wide range of possible objects and is thereby well suited for our purpose.

Time: 09.11.2010, 16:30-17:00
Location: Room 073 in the Fraunhofer IGD, Fraunhoferstrasse 5, S3|05
Speaker: Dominik Wodniok (Advisor: Jens Ackermann)
Title: "Realtime GPU-Raycasting of Volume Data with Smooth Splines on Tetrahedral Partitions" (Diploma thesis)
Abstract: (Talk will be held in German) The visualization of discrete volume datasets generated by imaging techniques like CT imaging and MRI has many important applications. Examples are medical diagnosis or quality assurance in industry. Common visualization methods are isosurfacing and full volume rendering. For both methods it is necessary to construct an at least continuous function from the discrete volume data. The quality of the images that are produced by these techniques strongly depends on the reconstruction method.

Two reconstruction methods proved to produce high quality visualizations of iso surfaces. They are trivariate splines in piecewise Bernstein-Bézier-form of polynomial degree two and three that are defined w.r.t. a tetrahedral partition. The current visualization approach with these splines exploits the massive computational power of GPUs. It combines rasterization of tetrahedra with precise per pixel raycasting. Realtime framerates were achieved for small datasets only. With increasing number of tetrahedra framerates become non interactive, as the sheer amount of geometry that has to be processed overburdens the GPU. Memory consumption becomes an issue as well. Full volume rendering has not been considered at all. We show, that the increased flexibility of modern GPUs allows to develop a geometry-free pure raycasting approach that makes realtime framerates feasible even for large datasets. Further, our approach allows the first implementation of interactive and realtime full volume rendering with these spline models.

Time: 09.11.2010, 17:00-17:30
Location: Room 073 in the Fraunhofer IGD, Fraunhoferstrasse 5, S3|05
Speaker: Dominik Wodniok (Advisor: Jens Ackermann)
Title: "Path Tracing in a Heterogeneous Multi-GPU Cluster" (Bachelor thesis)
Abstract: (Talk will be held in German) Path tracing has been an offline rendering technique ever since. With the enormous increase of speed seen with today's GPU hardware, interactive generation of images with global illumination effects becomes more and more feasible. In the near future we will probably see photorealistic computer graphic rendered in real time. While image footage was formerly created on render farms, the challenge today is distributing work across a network of machines equipped with GPUs.

In this work an approach to interactive global illumination is developed by assigning bidirectional path tracing workload to graphic cards in a cluster, making use of a framework for distributing CUDA jobs.

Time: 21.10.2010, 16:30-17:00
Location: Room 074 in the Fraunhofer IGD, Fraunhoferstrasse 5, S3|05
Speaker: Andreas Schwarzkopf (Betreuer Jens Ackermann)
Title: "GPU basierte volumetrische anisotrope Diffusion" (Diplomarbeit)
Abstract: (Talk will be held in German) Der Vortrag "GPU basierte volumetrische anisotrope Diffusion" beschäftigt sich mit der effizienten Umsetzung volumetrischer Diffusionsfilter auf modernen Grafikkarten und stellt die wesentlichen Überlegungen und Ergebnisse der gleichnamigen Diplomarbeit übersichtlich zusammen. Im Vordergrund stehen dabei die zu erwartenden Geschwindigkeitsvorteile, die sich durch eine geschickte GPU Implementierung und die damit einhergehende paralleliesierte Ausführung erzielen lassen. Vom physikalischen Hintergrund über die mathematische Formulierung bis hin zur Implementierung wird ein Überblick über die Thematik gegeben und die Ergebnisse des Prototypen diskutiert.

Time: 21.10.2010, 17:00-17:30
Location: Room 074 in the Fraunhofer IGD, Fraunhoferstrasse 5, S3|05
Speaker: Manuel Scholz (Betreuer Jens Ackermann)
Title: "Geometrieerfassung von Oberflächen mit einem analogen Lensshifter" (Diplomarbeit)
Abstract: (Talk will be held in German) Eine wichtige Aufgabe von VR Systemen ist die realistische Darstellung von Oberflächen. Dafür werden zunehmend komplexer werdende Modelle eingesetzt, die häufig auf einer Vermessung der Reflexionseigenschaften eines Materials beruhen. Jedoch sind auch geometrische Aspekte der Oberflächenstruktur ausschlaggebend für die Berechnung realistischer Bilder. Es existieren zwar eine Reihe von Algorithmen und Systemen zur Geometrievermessung jedoch sind diese meist nicht geeignet um feine Details ausreichend aufzulösen. Der Vortrag stellt die Entwicklung und Optimierung eines 3D-Rekonstruktionsverfahrens für einen 3D-Scanner vor, der unter Verwendung einer speziellen Projektorerweiterung den hohen Genauigkeitsanforderungen gerecht werden kann. Neben der Untersuchung und Vermessung der verwendeten Hardware wird speziell auf den Rekonstruktionsalgorithmus eingegangen.

Date: 30.09.2010, 16:30-17:00
Location: Room 074 in the Fraunhofer IGD, Fraunhoferstrasse 5, S3|05
Referent: Sebastian Lipponer (Advisor Ronny Klowsky)
Titel: "Point Based Rendering of Multi-View Stereo Depth Maps" (Diploma Thesis)

GRIS Colloquium

Date: 28.09.2010, 16:00
Location: Room 074 in the Fraunhofer IGD Fraunhoferstrasse 5, S3|05
Speaker: Prof. Nuno M. Guimaraes
University of Lisbon
Title: "Models for Implicit Human Environment Interaction"
Abstract: The design and construction of human-computer interfaces has been supported by conceptual and functional models that guided the reasoning approaches and the technical solutions for interactive system design and development. Central to this view has been the concept of mental model and a linguistic perspective of the interaction between the person and the computer.

The evolution of interactive computing, in particular the emergence of paradigms like ambient intelligence or ubiquitous or pervasive computing, poses meaningful challenges to the traditional conceptual and functional framework. Human-Environment Interaction, and especially implicit human-environment interaction, needs a new modeling framework.

Implicit human-environment interaction design should be based on the identification and construction of pattern-based relationships between the user state and the environment and break away from the tradition of the linguistic based models and architectures.

A set of of human environment interaction examples (some developed, some designed and under construction) like stress and anxiety management, reading support through physiological signals or driving fatigue detection, is presented. With this sample of cases, we aim to draw empirical evidence to support the new modelling and design principles.

Short Bio: Prof. Dr.-Ing. Nuno M. Guimares is a Professor at the Department of Informatics/Computer Science in the Faculty of Sciences of the University of Lisbon Portugal. His research domain is Human Computer Interaction and Interactive Systems Design with a current focus and interest on Ambient Intelligence and Ubiquitous Computing systems. In 2010 he was Invited Lecturer (Gastdozent) at the Technical University of Berlin (Institut fuer Psychologie und Arbeitswissenchaft), with the support of DAAD. He has been involved in EU-funded RD programmes since 1986 a co-founder of several technology-based companies in Portugal and Dean of the Faculty of Sciences from 2004 to 2009.

Date: 30.08.2010, 15:00
Location: Room 074 in the Fraunhofer IGD Fraunhoferstrasse 5, S3|05
Speaker: Prof. Wolfgang Heidrich
University of British Columbia
Title: "Visible Light Tomography in Computer Graphics"
Abstract: Tomographic methods are the standard approach for obtaining volumetric measurements in medicine, science, and engineering. Typical tomography setups acquire 2D X-ray images of an object, and reconstruct a 3D voxel representation from this data. Unfortunately, for many applications in computer graphics, such X-ray setups are not feasible due to cost and/or safety concerns.

In this presentation, I will introduce our recent work on visible light tomography, which has much more modest hardware requirements. I will discuss tomographic methods in the presence of refraction, and show applications to the scanning of transparent objects, and the capture of gas flows. I will also discuss first results on a new solver for generic tomography problems.

Short Bio: Professor Wolfgang Heidrich holds the Dolby Research Chair in Computer Science at the University of British Columbia. He received a PhD in Computer Science from the University of Erlangen in 1999, and then worked as a Research Associate in the Computer Graphics Group of the Max-Planck-Institute for Computer Science in Saarbrucken, Germany, before joining UBC in 2000. Heidrich's research interests lie at the intersection of computer graphics, computer vision, imaging, and optics. In particular, he has worked on High Dynamic Range imaging and display, image-based modeling, measuring, and rendering, geometry acquisition, GPU-based rendering, and global illumination. Heidrich has written over 100 refereed publications on these subjects and has served on numerous program committees. He was the program co-chair for Graphics Hardware 2002, Graphics Interface 2004, and the Eurographics Symposium on Rendering, 2006.

Invited Talk

Time: 18.08.2010, 13:30
Location: Room 074 in the Fraunhofer IGD Fraunhoferstrasse 5, S3|05
Speaker: Prof. Dr. Daniel Cremers
TU München
Title: "Global Optimization Methods for Computer Vision"
Abstract: A multitude of computer vision challenges can be cast as problems of energy minimization. One of the major scientific challenges lies in efficiently computing solutions of minimal energy. I will introduce optimization methods which address problems such as the segmentation of moving objects in image sequences, the detection of obstacles in traffic videos, or the reconstruction of 3D shapes from a collection of 2D images. I will detail how respective cost functionals can be minimized both by continuous (PDE and level set methods) and by discrete (graph theoretic) methods.

In particular, I will show that the classical problem of multiview reconstruction can be solved by continuous convex relaxation methods. In contrast to most existing energy minimization methods, the proposed algorithms do not require an initialization and allow to compute robust solutions with guaranteed optimality properties.

GRIS Colloquium

Date: 03.08.2010, 16:30
Location: Room 074 in the Fraunhofer IGD Fraunhoferstrasse 5, S3|05
Speaker: Prof. Dr. Benjamin Bustos
University of Chile
Title: "On Index-free Similarity Search in Metric Spaces"
Abstract: Metric access methods (MAMs) serve as a tool for speeding up similarity queries. However, all MAMs developed so far are index-based; they need to build an index on a given database. The indexing itself is either static (the whole database is indexed at once) or dynamic (insertions/deletions are supported), but there is always a preprocessing step needed.

In this talk I will present the D-file, the first MAM that requires no indexing at all. This feature is especially beneficial in domains like data mining, streaming databases, etc., where the production of data is much more intensive than querying. Thus, in such environments the indexing is the bottleneck of the entire production/querying scheme. The idea of D-file is an extension of the trivial sequential file (an abstraction over the original database, actually) by the so-called D-cache. The D-cache is a main-memory structure that keeps track of distance computations spent by processing all similarity queries so far (within a runtime session). Based on the distances stored in D-cache, the D-file can cheaply determine lower bounds of some distances while the distances alone have not to be explicitly computed, which results in faster queries. An experimental evaluation shows that query efficiency of D-file is comparable to the index-based state-of-the-art MAMs, however, for zero indexing costs.

Short Bio: Benjamin Bustos is Assistant Professor in the Department of Computer Science at the University of Chile. He is head of the PRISMA (Pattern Recognition, Indexing, and Similarity search in Multimedia Archives) Research Group. He leads several research projects in the domains of multimedia retrieval, video copy detection, sketch-based image retrieval, and automatic processing of handwritten documents. Recently, he has been awarded by Yahoo! Labs to further his project on Web Multimedia Retrieval. His research interests include similarity search, multimedia information retrieval, 3D object retrieval, and indexing in (non)-metric spaces. He obtained a doctoral degree in natural sciences from the University of Konstanz, Germany, in 2006.

Date: 14.07.2010, 13:30
Location: Room 074 in the Fraunhofer IGD Fraunhoferstrasse 5, S3|05
Speaker: Prof. Dr. Klaus-Robert Mueller
TU Berlin, Machine Learning Group
Bernstein Center for Computational Neuroscience
Quantitative Products Lab
Title: "Machine Learning and Applications"
Abstract: This lecture provides a brief introduction to Support Vector Machines as an example for successful kernel-based machine learning (ML) and touches on fundamentally open issues in this field. Then a short overview of selected successful applications of kernel based ML is given, covering e.g. hacker intrusion detection, computer vision and biomedical engineering (EEG-based Brain Computer Interfaces).

Invited Talk

Date: 29.6.2010, 10:00
Location: Room 074 in the Fraunhofer IGD Fraunhoferstrasse 5, S3|05
Speaker: Prof. van Wijk
TU Eindhoven
Title: "Knots, Maps, and Tiles: Mathematical Visualization Puzzles"
Abstract: Visualization can help to get insight into mathematical objects, and how to obtain a suitable visualization can be an intriguing puzzle. Three such puzzles are discussed. Seifert surfaces are orientable surfaces bounded by mathematical knots. How to construct geometric models of these? Mapping the earth is a classic problem. If we allow for interrupts, almost distortion free mappings can be obtained. What are the possiblities here? Finally, a regular map is a symmetric tesselation of a closed surface. How to construct a space model of these objects? All cases are illustrated with many images and animations.

GRIS Colloquium

Date: 11.05.2010, 16:30
Location: Room 074 in the Fraunhofer IGD Fraunhoferstrasse 5, S3|05
Speaker: Prof. Jan Bender
TU Darmstadt
Title: "Impulsbasierte Dynamiksimulation"
Abstract: (Talk will be held in German) Die dynamische Simulation ist ein aktueller Forschungsbereich mit vielen Anwendungsgebieten. Dieser Bereich umfasst u.a. die Simulation von Starrkörpern, Textilien und Weichkörpern. Dynamische Simulation ist heutzutage ein wichtiger Bestandteil im Entwicklungsprozess von komplexen Maschinen (wie z.B. in der Robotik), im Forschungsbereich der Medizin, in der Computeranimation und der virtuellen Realität.

An der Universität Karlsruhe wurde in den letzten Jahren ein neues Verfahren für die Simulation von Starrkörpern entwickelt, die durch Gelenke miteinander verbunden sind. Jedes Gelenk definiert eine Zwangsbedingung für die verbundenen Körper. Diese Bedingung wird im Gegensatz zu klassischen Verfahren mit Hilfe von Impulsen aufgelöst. Dadurch kann die Simulation effizient durchgeführt werden und liefert sehr genaue Ergebnisse. Ein weiterer Vorteil der impulsbasierten Simulation ist, dass Kollisionen und bleibende Kontakte mit Reibung auf die gleiche Art wie Gelenke simuliert werden können und sich dadurch ein einheitliches Verfahren ergibt.

Für den impulsbasierten Ansatz wurde mittlerweile ein Algorithmus entwickelt, der die Simulation für azyklische Modelle mit dem optimalen Zeit- und Speicheraufwand von O(n) ermöglicht. Dadurch können selbst sehr komplexe Modelle in Echtzeit simuliert werden. Außerdem konnte die impulsbasierte Simulation durch die Parallelisierung eines iterativen Lösungsverfahrens sehr effizient auf einem Graphikprozessor implementiert werden. Das ursprüngliche Verfahren wurde für die Simulation von Textilien und Weichkörpern erweitert. Die Simulation von Textilien wird hauptsächlich in der Computeranimation und in Computerspielen angewendet, während die Simulation von Weichkörpern auch im medizinischen Bereich von großem Interesse ist.

Thesis presentations

Date: 29.04.2010, 10:45-11:30
Location: Room 048 in the Fraunhofer IGD, Fraunhoferstrasse 5, S3|05
Speaker: Azam Sadat Ghassemi Hosseini (Advisor Matthias Kirschner)
Title: (Talk will be held in German) "Knochenmineralbestimmung mittels Dual-Energy CT" (Diploma Thesis)
Abstract: Osteoporose ist eine systemische Skeletterkrankung, die durch eine Verminderung der Knochenmasse und Veränderungen in der Knochenmikrostruktur charakterisiert ist. Das in dieser Diplomarbeit vorgestellte Verfahren basiert auf einem Knochendichtemessverfahren, das eine Darstellung der Knochendichte aus entsprechenden dreidimensionalen Dual-Energy-CT-Datensätzen ermöglicht. In einem ersten Schritt wird die spongiose Knochensubstanz in jeder einzelnen CT-Schicht identifiziert und markiert. Anschließend wird die Knochenmineraldichte und die Dichte des trabekulären Raums auf Basis der CT-Zahlen der Datensätze berechnet. Abschließend werden diese Werte den Voxeln der CT-Aufnahmen der Wirbelsäule für eine interaktive Untersuchung der Knochendichte im trabekulären Knochen überlagert.

Date: 29.04.2010, 10:00-10:45
Location: Room 048 in the Fraunhofer IGD, Fraunhoferstrasse 5, S3|05
Speaker: Ferry Wibowo (Advisor Kawa Nazemi)
Title: "Visualization of Time-Dependent Semantics Data" (Bachelor Thesis)
Abstract: Time is an essential notion and common attribute in information science. Through the addition of the semantics structure on that information, it can turn out to be a powerful fundamental data in which the knowledge within can be imparted. Nevertheless, this high potential capability is the contrary side that makes the development of adequate graphical interface even more arduously. Timeline is commonly used for addressing linear time-dependent data and applied as an excellent instrument to represent it. However, most timeline collide with the matter when they encounter a set of complex data, such as the semantics data. Following this further, this thesis proposes a new conceptual design to tackle the hindrances. It supports rich visualization by permitting dynamic hierarchy, cross concept relationships, meta-relationship, fast overview, scalability, and trend recognition. Moreover, the visualization concept is also geared with a set of interactions features which take the advantages of semantic technology to its fullest. It allows exploring the data through the time, distorting time, and generally brushes the data.

Date: 19.04.2010, 14:00-14:30
Location: Room 048 in the Fraunhofer IGD, Fraunhoferstrasse 5, S3|05
Speaker: Tobias Bauer (Advisor Matthias Kirschner)
Title: (Talk will be held in German) "Parallele Ansätze für Active Shape Models" (Diploma Thesis)
Abstract: In der heutigen Medizin wird verstärkt eine Segmentierung medizinischer Bilder zur Diagnose und Operationsplanung eingesetzt. Eine genaue und schnelle Segmentierung ist dabei ein komplexer Vorgang. Kenntnisse über Struktur und Form der zu segmentierenden Bereiche vereinfachen den Vorgang der Segmentierung. Ein Algorithmus, der diese Art Vorwissen berücksichtigt, ist der Active-Shape-Model-Algorithmus (ASM). Diese Diplomarbeit befasst sich mit parallelen Ansätzen des ASM-Algorithmus. Im Rahmen der Arbeit wurden verschiedene Ansätze des Matchingschrittes im ASM-Algorithmus in serieller und paralleler Ausführung evaluiert. Für die Parallelisierung wurde OpenMP und CUDA eingesetzt. Durch den Einsatz paralleler Architekturen konnten einzelne Rechenschritte um den Faktor 10, der gesamte ASM-Algorithmus um den Faktor 2-4 beschleunigt werden. Neben der Beschleunigung wurde die Segmentierungsgenauigkeit untersucht. Dabei zeigten Matchingstrategien, die auf statistischer Modellierung der lokalen Struktur aufbauen, die besten Ergebnisse.

Date: 16.04.2010, 09:30-10:00
Location: Room 011 in the Fraunhofer IGD, Fraunhoferstrasse 5, S3|05
Speaker: Matthias Noll (Advisor Matthias Keil)
Title: (Talk will be held in German) "Navigation für minimalinvasive Eingriffe an der Niere unter Verwendung von Augmented Reality Elementen" (Diploma Thesis)

Date: 16.04.2010, 10:00-10:30
Location: Room 324 in the Fraunhofer IGD, Fraunhoferstrasse 5, S3|05
Speaker: Ardalan Naseri and Joseph Karuiru (Advisor Tatiana von Landesberger)
Title: "Graph visualization with focus on graph layouts" (Practical Work)

Date: 13.04.2010, 16:30-17:00
Location: Room 048 in the Fraunhofer IGD, Fraunhoferstrasse 5, S3|05
Speaker: Robert Rehner (Advisor Tatiana von Landesberger)
Title: (Talk will be held in German) "Verbesserung der Performanz von Graphanalyse und Visualsierung vor - mit einem Focus auf Motif-Suche im RelaNet-Tool" (Practical Work)

Date: 13.04.2010, 17:00-17:30
Location: Room 048 in the Fraunhofer IGD, Fraunhoferstrasse 5, S3|05
Speaker: Phillip Hartmann (Advisor Tatiana von Landesberger)
Title: (Talk will be held in German) "Portierung des Coherent Point Drift-Algorithmus zur nicht-rigiden Registrierung von Punktwolken von Matlab nach C++" (Practial Work)

Date: 01.03.2010, 16:00-16:30
Location: Room 048 in the Fraunhofer IGD, Fraunhoferstrasse 5, S3|05
Speaker: Tarik Tahiri, Waheed Abdul, Youssef Bouchiba (Advisor Maximilian Scherer)
Title: "Open Art Library" (Practical Work)

Date: 28.01.2010, 16:30-17:00
Location: Room 048 in the Fraunhofer IGD, Fraunhoferstrasse 5, S3|05
Speaker: Alexander Marinc (Advisor Jens Ackermann)
Title: "Interaktive Visualisierung von C1-stetig differenzierbaren quadratischen Splines auf abgeflachten Oktaederpartitionen" (Master Thesis)
Abstract: (Talk will be held in German) Die Rekonstruktion diskreter Daten findet in vielen Gebieten Anwendung. Egal ob wissenschaftliche Strömungsanalysen oder medizinische CRT-Daten, es wird eine möglichst genaue und kontinuierliche Approximation der Ursprungsdaten benötigt, um die Daten effizient auswerten zu können. Gerade in Bezug auf die Visualisierung der impliziten Oberflächen ist eine C1-stetige Rekonstruktion wünschenswert, um auch unter Beleuchtung ansprechende und glatte Ergebnisse zu erhalten. Trivariate BB-Splines wurden bereits erfolgreich auf Typ-6-Tetraeder-Partitionen eingesetzt, um hochqualitative Visualisierungen umzusetzen, welche Renderíng und Isowertwechsel in Echtzeit zulassen. Bisher wurde jedoch ein kubischer Grad der einzelnen Patches benötigt, um echte C1-Stetigkeit im gesamten Volumen ermöglichen zu können.

In dieser Arbeit wird zum ersten Mal eine Partition visualisiert, welche mit quadratischem Grad diese Bedingung erfüllen kann. Sie basiert auf abgeflachten Oktaedern, welche in jeweils 144 Tetraeder geteilt werden. Bereits in einer vorherigen Arbeit wurden die besseren approximativen Eigenschaften der neuen Partition nachgewiesen, sowie analytische und numerische Beweise für ihre Korrektheit erbracht. Diese Ergebnisse werden in dieser Arbeit visuell bestätigt und verwendet, um die Isooberflächen verschiedenster Volumendaten effizient mit GPU unterstütztem Raycasting anzuzeigen. Viele der bereits für die Typ-6-Splines erprobten Techniken kommen hier zum Zuge und ermöglichen durch GPU-basierte Algorithmen, trotz der deutlich erhöhten Komplexität der neuen Partition, ebenfalls Berechnungen in kürzesten Zeitspannen. Neben der erfolgreichen Erprobung und Verbesserung der bisherigen Mittel werden auch neue Aspekte, wie eine Verlagerung der Koeffizientenberechnung in der programmierbaren Graphikpipeline und die Möglichkeit mehrere Isooberflächen anzeigen zu können, mit eingebracht. Auch der Vergleich mit den Typ-6-Splines in Bezug auf die Berechnungs- und Visualisierungszeiten, sowie eine qualitative Analyse aller Ergebnisse mit dem Oktaederspline stellt einen der Hauptpunkte in dieser Arbeit dar. Wir werden zeigen, dass sich mit dem Oktaederspline effiziente und hochqualitative visuelle Ergebnisse erzielen lassen, welche modernen Standards in jeder Hinsicht gerecht werden und zudem sehr gute Rekonstruktionseigenschaften beinhalten..

GRIS Colloquium

Date: 29.01.2010, 13:30
Location: Room 074 in the Fraunhofer IGD Fraunhoferstrasse 5, S3|05
Speaker: Prof. Carsten Dachsbacher
Universitaet Stuttgart
Title: "Analyzing Visibility Configurations"
Abstract: Visibility determination has been a central problem in computer graphics since the very beginning of the field. It is required for many aspects such as the detection of visible and occluded surfaces to speed up image generation, computing shading and shadowing from light sources, the simulation of global illumination, and even in acoustics simulation.

In interactive rendering applications visibility is typically used to prevent invisible or occluded geometry from being processed. This talk addresses the question how this visibility information can be used beyond the culling of occluded geometry. A novel method for classifying “visibility configurations”, takes the distribution, or structure, of visibility across surfaces into account. The possibility to identify and differentiate between these configurations provides valuable information for many algorithms relying on visibility determination. This method can be combined with findings from perceptual rendering methods and initial results demonstrate promising applications in real-time and off-line rendering.

Thesis presentations

Date: 25.01.2010, 16:00-16:30
Location: Room 074 in the Fraunhofer IGD, Fraunhoferstrasse 5, S3|05
Speaker: Markus Schnoes (Advisor Jens Ackermann)
Title: "Massively Parallel SVBRDF Fitting of BTF Data" (Bachelor Thesis)
Abstract: In order to generate realistic looking images, modern methods for global illumination require detailed material representations. With the Bidirectional Texture Function (BTF) an image-based technology exists that fulfills these requirements at the expense of a high memory consumption and difficult editability. By separating a BTF in its heightfield and an analytical spatially varying Bidirectional Reflectance Distribution Function (SVBRDF) an efficient compression and good editability can be achieved, while maintaining the essential reflection characteristics.

In this thesis a hybrid method using both CPU and GPU is presented to generate analytical SVBRDFs on the base of a BTF and the according heightfield. The focus is put upon an efficient work balance between both processing units, in order to speed up the fitting process significantly. For this purpose an appropriate analytical SVBRDF model has been developed. Additionally this thesis describes a hybrid, parallel working optimization process and evaluates it with the help of selected BTF datasets. Finally the results are discussed in detail. In comparison to conventional Methods, the approach presented in this work achieves a significant speedup in the optimization phase. The resulting representation consisting of SVBRDF and heightfield can additionally, in contrast to BTF data, be integrated easily in a shader-based rendering algorithm while still maintaining real-time capabilities.

Date: 18.01.2010, 15:30-16:00 Uhr
Location: Room 074 in the Fraunhofer IGD, Fraunhoferstrasse 5, S3|05
Speaker: Michael Schmitt (Advisor Stefan Wesarg)
Title: "Evaluation statistischer Verfahren zur Modellierung von Formvariation" (Bachelor Thesis)
Abstract: /Talk will be in german) Statistische Modelle werden in vielen Bereichen der medizinischen Bildverarbeitung eingesetzt, aber auch zur Lösung von allgemeinen Aufgaben in Computer Vision und Computer Graphik, wie etwa bei der Gesichtserkennung, der Entrauschung von Bildern oder der Segmentierung. In der Theorie wurden mehrere statistische Methoden zur Verwendung bei den genannten Problemen entwickelt, die aber im praktischen Einsatz nur geringen Anklang finden. So wird im Kontext der statistischen Formmodelle größtenteils die Hauptkomponentenanalyse angewandt. In dieser Arbeit untersuche ich statistische Modelle, die im Rahmen des Active Shape Models zur Bildsegmentierung eingesetzt werden können. Das Ziel der Arbeit ist es, durch diese Untersuchung herauszufinden, worin die Diskrepanz zwischen Theorie und Praxis begründet ist. Ich suche also nach Ursachen dafür, dass zwar verschiedenste Techniken bekannt, in der Wirklichkeit aber nur die Hauptkomponentenanalyse eingesetzt wird. Dafür wurden zwei weitere wichtige Vertreter statistischer Modelle ausgewählt, welche im Folgenden vorgestellt und anhand ihrer Leistungsfähigkeit miteinander und mit der Hauptkomponentenanalyse verglichen werden.

Date: 18.01.2010, 16:15-16:45
Location: Room 074 im Fraunhofer IGD, Fraunhoferstrasse 5, S3|05
Speaker: Florian Jung (Advisor Stefan Wesarg)
Titlel: "GPU-basierte rigide Registrierung von 3D-Bilddaten" (Bachelor Thesis)
Abstract: /Talk will be in german) Diese Bachelorarbeit beschäftigt sich mit der Frage, inwieweit sich der Prozess der rigiden Bildregistrierung auf der Grafikkarte implementieren lässt. Hierfür wird Nvidias CUDA verwendet. Zuerst wird der Algorithmus auf Parallelisierbarkeit untersucht und anschließend auf die Grafikkarte portiert, um so eine Beschleunigung gegenüber der CPU-Implementierung zu erreichen. Anschließend wird die Geschwindigkeit und Genauigkeit der beiden Implementierungen verglichen und auf unterschiedlichen Systemen getestet. Durch die Verwendung von CUDA konnte der Algorithmus um den Faktor 4-10 beschleunigt werden, abhängig von der verwendeten Hardware.

GRIS Colloquium

Date: 14.01.2010, 13:00
Location: Room 074 in the Fraunhofer IGD Fraunhoferstrasse 5, S3|05
Speaker: Prof. Fernando De la Torre
Carnegie Mellon University
Title: "Learning Components for Human Sensing"
Abstract: Providing computers with the ability to understand human behavior from sensory data (e.g. video, audio, or wearable sensors) is an essential part of many applications that can benefit society such as clinical diagnosis, human computer interaction, and social robotics. A critical element in the design of any behavioral sensing system is to find a good representation of the data for encoding, segmenting, classifying and predicting subtle human behavior. In this talk I will propose several extensions of Component Analysis (CA) techniques (e.g. kernel principal component analysis, support vector machines, and spectral clustering) that are able to learn spatio-temporal representations or components useful in many human sensing tasks.

In the first part of the talk I will give an overview of several ongoing projects in the CMU Human Sensing Laboratory, including our current work on depression assessment and deception detection from video, as well as hot-flash detection from wearable sensors. In the second part of the talk I will show how several extensions of the CA methods outperform state-of-the-art algorithms in problems such as temporal alignment of human behavior, temporal segmentation/clustering of human activities, joint segmentation and classification of human behavior, and facial feature detection in images. The talk will be adaptive, and I will discuss the topics of major interest to the audience.

Thesis presentations

Date: 16.12.2009, 09:00-09:30
Location: Room 242 in the Fraunhofer IGD, Fraunhoferstrasse 5, S3|05
Speaker: Christoph Kreitz (Advisor Klaus Drechsler)
Title: "Segmentvisualisierung der Leber für die intraoperative Navigation" (Diploma thesis)
Abstract: (Talk will be held in German) Für einige Lebertumore ist die vollständige chirurgische Entfernung die derzeit einzige potentiell heilende Therapie. Die Operabilität ist von diversen Faktoren abhängig: 1. Der Tumor muß mit ausreichend Sicherheitsabstand zum gesunden Gewebe entfernt werden. 2. Das restliche Lebergewebe muß eine ausreichende Funktionsleistung erbringen, welche vom verbleibenden Lebervolumen abhängt. 3. Von der Durchblutung abgetrennte Bereiche müssen erkannt und ebenfalls entfernt werden. Drei Operationsstrategien berücksichtigen diese Faktoren: 1. Hemihepatektomie (komplette Entfernung einer Leberhälfte). 2. Tumororientierte Operation (Keilförmiges Stück der Leber wird entfernt). 3. Segmentorientierte Operation (Eins oder mehrere der Lebersegmente wird entfernt). Zur unterstützenden Operationsplanung und zur besseren Orientierung des Chirurgen während des Eingriffs wurde in dieser Arbeit ein Software-Tool entwickelt, welches patientenspezifische Leber(sub)segmente aus präoperative CT Daten mit möglichst wenig User-Interaktion anhand von Abstraktionen der Lebergefäße berechnet und visualisiert.

GRIS Colloquium

Date: 11.12.2009, 13:00
Location: Room 074 in the Fraunhofer IGD Fraunhoferstrasse 5, S3|05
Speaker: Prof. David J. Fleet
Department of Computer Science (University of Toronto)
Title: "Physics-Based Models for Human Motion Analysis"
Abstract: The recovery and analysis of human motion from video is a key enabling technology for myriad applications (e.g., in man-machine interaction, computer graphics, biometrics, and biomechanics). Many are confident that the problem will be solved, in part with the help of models for how people move. Current state-of-the-art models are usually learned from human motion capture data, but there are questions about whether such models will work well in unconstrained situations.

In this talk we advocate a new class of models, derived in part from principles of Newtonian dynamics and biomechanics. We describe two examples of such models: The first, inspired by low-dimensional passive-dynamic models of human locomotion, was designed for monocular tracking of walking people. The second uses physical principles to facilitate the inference of human muscle forces and ones interactions with external surfaces.

Joint work with Marcus Brubaker and Leon Sigal.

Thesis presentations

Date: 08.12.2009, 14:45-15:15
Location: Room 324 in the Fraunhofer IGD, Fraunhoferstrasse 5, S3|05
Speaker: Jürgen Bernard (Advisor Tobias Schreck, Tatiana von Landesberger)
Title: "Methoden zur Qualitätsbewertung von Self-Oranizing Maps zur Unterstützung des visuellen Analyseprozesses" (Diploma thesis)
Abstract: (Talk will be held in German) Der Self-organizing Maps Algorithmus (SOM) erfreut sich in der visuellen Clusteranalyse großer Beliebtheit. Hauptgründe hierfür sind seine topologieerhaltende Eigenschaft bei der Partitionierung von hochdimensionalen Datensätzen und die leichte Visualisierbarkeit von Ergebnissen aufgrund der SOMGitterstruktur. Seit ihrer Einführung, wurden SOMs in einer Vielzahl unterschiedlicher Data Mining Anwendungen verwendet. Einhergehend mit dem technischen Fortschritt von Computersystemen, haben sich die Visualisierungs- und Interaktionstechniken in der visuellen Clusteranalyse kontinuierlich verbessert. Maßgeblichen Anteil an dieser Entwicklung haben die Fachgebiete der Informationsvisualisierung und die aufstrebende Disziplin der Visual Analytics. In jüngster Zeit wurden zukunftsweisende, visuell-interaktive Repräsentierungen des SOM-Algorithmus zur Informationsbeschaffung aus komplexen Datensätzen vorgestellt. Ein aktueller Trend der Visual Analytics ist die visuell-interaktive Erschließung des gesamten Clusteranalyseprozesses, beginnend mit der Datenvorverarbeitung, bis hin zur Evaluierung von Clusteringergebnissen. Dies schließt auch die schrittweise Verbesserung der erreichten Clusteringqualität mit ein (Iterative Refinement). Die durch gesteigerte Einflussmöglichkeiten des Menschen vollzogene Distanzierung von vollautomatischen Clusteranalysetools, birgt enormes Potential bei der Exploration von Daten. In gleicher Weise gilt es jedoch, potentielle Schwierigkeiten neuartiger Clusteranalysetools zu berücksichtigen. Besondere Aufmerksamkeit bedarf es in diesem Zusammenhang der Bewertung und Visualisierung der erreichten SOM-Clusteringqualität, die durch erweiterte Möglichkeiten der Einflussnahme maßgeblich vom Anwender abhängig ist. Die Notwendigkeit von Werkzeugen zur visuellen Qualitätsbewertung erstreckt sich hierbei vom Prozessschritt des Iterative Refinements bis weit in die Postprocessingphase des Clusterings hinein. Das Ziel dieser Diplomarbeit besteht in der Entwicklung von Lösungsstrategien um den SOMClusteranalyseprozess mit visuellen Qualitätsbewertungsmethoden zu unterstützen. Umgesetzt werden die Ergebnisse im FinExplorer Framework, einer SOM-basierten Applikation zur visuellen Analyse von Finanzdaten, die um entsprechende Methoden zur visuellen Qualitätsbewertung erweitert wird. Neben der Integration bewährter Darstellungsformen, liegt der Schwerpunk dieser Arbeit in der Konzeption, der Umsetzung und der Evaluation neu entwickelter Visualisierungsformen. Der Umfang der Arbeit erstreckt sich dabei über die Erarbeitung von trainingsbegleitenden Qualitätsvisualisierungen, die Erschließung und Visualisierung von geeigneten statistischen Qualitätsindices, die verbesserte Einsicht auf die SOM-Struktur mit ihrem zugrundeliegenden Datensatz und schließlich auf die komparative Darstellung der SOM mit Ergebnissen aus anderen Clusteringalgorithmen.

Date: 08.12.2009, 15:30-16:00
Location: Room 324 in the Fraunhofer IGD, Fraunhoferstrasse 5, S3|05
Speaker: Sebastian Fiebig (Advisor Tatiana von Landesberger)
Title: "Interaktionstechniken und -tracking in der Visualisierung von Graphen" (Bachelor thesis)
Abstract: (Talk will be held in German) The analysis of corporate structures has high importance for the financial sector. Corporate structures consist of companies holding shares in other companies.
For analyzing corporate structures a graphical visualization of these networks is an effective instrument. An efficient analysis needs adequate interaction techniques, which for example allow restrict the amount of visualized data. This can be achieved by filtering the data or highlighting of structural motifs in the graphs representing the corporate structures. It is significant for the user's work to visualize her sequence of interactions with software systems and analyze them. This enables the user to understand, reproduce and optimize her work process, resulting in an optimization in usage of the software and optimization of the software itself. This bachelor thesis extends RelaNet, a program for visualizing graphs with specialized functions for analyzing corporate structures, by implementing interaction techniques, like filtering, highlighting and aggregation, based on a search for vertex and edge attributes. The user achieves a more tightly focused and easier analysis, because of a more focused visualization of data. Furthermore a tracking will be implemented, which saves user actions, beyond one session, visualize them, allows navigation between program states and thus enables an analysis of the user's analytical process.

With the help of a case study the functionality and advantages of the presented implementation are shown. The case study uses a database containing 12 000 German corporations. The case study shows how a better analysis is possible by using the new interaction techniques in combination with the tracking system. The usage in other information visualization software such as for biology or supply chain management is also possible.

GRIS Colloquium

Date: 04.12.2009, 16:00
Location: Room 074 in the Fraunhofer IGD Fraunhoferstrasse 5, S3|05
Speaker: Prof. Andries van Dam
Computer Graphics Group (Brown University)
Title: "Fluid Interaction with Pen- and Touch-Computing"
Abstract: The great popularity of the iPhone and its touch-enabled competitors has brought the era of post-WIMP UIs to consumers and has made flick and pinch gestures as ubiquitous as point-and-click became after the introduction of the Macintosh. Post-WIMP UIs, such as pen- and touch-computing and gesture-recognition for the specification of both content and commands, are actually pre-WIMP since they date from the mid-1960s when digitizer tablets such as the Rand Tablet first appeared. In that same era Doug Engelbart pioneered bi-manual operation of interaction devices such as the mouse and chording keyboard in his landmark NLS system. There were also precedents for pen- and touch-computing in the consumer space such as the Apple Newton, the Graffiti-driven Palm PDA, and tablet PCs.

Today there is an explosion of interest both in research and in new consumer products based on multi-touch interaction techniques, using a variety of form factors ranging from cell phones through small tablets to large tabletop displays such as the Microsoft Surface. In this talk I will show demos of some of my group's work on both pen- and touch-computing and talk about current research challenges such as how to take better advantage of bi-manual operation, how to expand our gesture vocabulary, and means for discovering and learning gestures.

Date: 24.11.2009, 16:30
Location: Room 074 in the Fraunhofer IGD Fraunhoferstrasse 5, S3|05
Speaker: Torsten Ullrich
Geschäftsbereich Visual Computing (Fraunhofer Austria Research)
Title: "Semantic Fitting and Reconstruction"
Abstract: Generative modeling techniques have influenced the modeling process significantly. Rule-based models such as cities and urban landscapes can be created in arbitrary size without much effort. The resulting models may appear realistic, but are not the reconstruction of some real objects and their geometry. The process to bring the geometry of a real object together with a suitable shape template and to extract its main characteristic parameters is known as shape recognition and reverse engineering. Reverse engineering of generative models offers a new advantage: semantic enrichment. With more and more virtual objects in model repositories algorithms gain importance, which are able to recognize a shape, to extract its main parameters and therefore to classify a model and to enrich it semantically.

Thesis presentations

Date: 19.10.2009, 11:15-11:45
Location: Room 324 in the Fraunhofer IGD, Fraunhoferstrasse 5, S3|05
Speaker: Ralf Gutbell (Advisor Stefan Wesarg)
Title: "Fokussierte Visualisierung medizinischer Bilddaten" (Bachelor thesis)

Date: 17.11.2009, 16:30-17:00
Location: Room 074 in the Fraunhofer IGD, Fraunhoferstrasse 5, S3|05
Speaker: Simon Fuhrmann (Advisor Jens Ackermann)
Title: "Curvature-adaptive, Feature-sensitive Isotropic Surface Remeshing" (Master thesis)
Abstract: (Talk will be held in German) The issue of surface remeshing aims at creating a new triangular discretization for a given input mesh with respect to certain quality characteristics, i.e. the task is to improve mesh quality in some sense. The need for remeshing is induced by the fact that raw meshes, e.g. those resulting from an automated scanning process, are rarely suited for further processing. This also holds for typical CAD models with a lot of degenerated triangles, which makes it almost impossible to get reliable simulation results for these meshes without reparation. The discipline of remeshing has a long history and an abundance of techniques has been proposed over time. However, the recent success of relaxation-based methods shows a clear trend in high-quality mesh generation. In this work a general framework for relaxation-based remeshing is presented, applicable to a wide range of models and published as free application for the scientific community.

GRIS Colloquium

Date: 12.11.2009, 16:30
Location: Room 074 in the Fraunhofer IGD Fraunhoferstrasse 5, S3|05
Speaker: Univ.-Prof. Dr. Oliver Bimber
Institute of Computer Graphics
(Johannes Kepler University Linz)
Title: "Projection over Four Orders of Magnitude"
Abstract: The combination of real-time visualization and image analysis with adapted imaging and display optics enable a large variety of new applications for projector-camera systems. These are ranging from intelligent illumination techniques in radiology, endoscopy and microscopy, over digital video composition techniques for visual effects in TV studios and non-studio film-sets, to video projections on everyday surfaces in museums or historic sites. In this talk, I will give an overview over techniques and results from recent projects, such as SmartProjecting, VirtualStudio2Go, Superimposing Dynamic Range, Temporal Backdrops, Coded Aperture Projection and Projected Light Microscopy. I will show how pixel size and contrast can vary over four orders of magnitude for different display and imaging applications that utilize projector-camera systems.

Date: 03.11.2009, 16:30
Location: Room 073 in the Fraunhofer IGD Fraunhoferstrasse 5, S3|05
Speaker: Tom Haber
Expertise centre for Digital Media (Universiteit Hasselt)
Title: "Relighting objects from image collections"
Abstract: Applications in entertainment, augmented/virtual reality, architecture and digital museums require the creation of realistic computer-generated images of real-world objects from novel viewpoints and with novel illumination. These images should be indistinguishable from real photographs taken under similar conditions. For this purpose, we will look at an approach for recovering the reflectance of a static scene with known geometry from a collection of images taken under distant, unknown illumination. Using an all-frequency relighting framework based on wavelets, we are able to simultaneously estimate the per-image incident illumination and the per-surface point reflectance. The wavelet framework also allows for incorporating various reflection models. Combined with multi-view stereo reconstruction, we are even able to recover the geometry and reflectance of a scene solely using images collected from the Internet.

Thesis presentations

Date: 15.10.2009, 10:00-10:30
Location: Room 074 in the Fraunhofer IGD, Fraunhoferstrasse 5, S3|05
Speaker: Martin Ritz (Advisor Jens Ackermann)
Title: "A Combined Multi-View Stereo and Photometric Stereo Approach" (Master thesis)
Abstract: Two approaches to 3D-reconstruction from images, Multi-view Stereo and Photometric Stereo, are combined as a means of completing missing areas within the partly-complete reconstruction returned by the first approach. This is realized by exploiting the Photometric Stereo principle of orientation conistency, stating that image positions with similar appearance have similar surface orientation on the original object surface. Completion of recontruction is thus achieved by finding pairs of image positions with similar appearance, one with available reconstruction serving as source, the other one with recontruction missing as target for transfer of normals.

Date: 05.10.2009, 11:30-12:00
Location: Room 074 in the Fraunhofer IGD, Fraunhoferstrasse 5, S3|05
Speaker: Anna Mitkova (Advisor Tatiana von Landesberger)
Title: "Methoden für die Visualisierung von Zeitabhängigen Graphen" (in german, Bachelor thesis)

Date: 05.10.2009, 12:00-12:30
Location: Room 074 in the Fraunhofer IGD, Fraunhoferstrasse 5, S3|05
Speaker: Vyara Ivanova (Advisor Tatiana von Landesberger)
Title: "Theoretische Grundlagen zur Visualisierung und Analyse von Netzwerken und Hierarchien bezogen auf die Aktionärs- und Branchenstruktur von Unternehmen" (in german, Diploma thesis)

Date: 09.09.2009, 15:30-16:00
Location: Room 073 in the Fraunhofer IGD, Fraunhoferstrasse 5, S3|05
Speaker: Michael Walter (Advisor Tobias Schreck)
Title: "Implementierung und Evaluierung einer gradientbasierten Methode für das 3D Model Retrieval" (in german, Bachelor thesis)

Date: 02.09.2009, 15:00-15:30
Location: Room 074 in the Fraunhofer IGD, Fraunhoferstrasse 5, S3|05
Speaker: Sebastian Maier (Advisor Sebastian Bremm)
Title: "Entwicklung und Bewertung eines Ansatzes zur visuellen Analyse in zeitabhängigen Sequenzdaten" (in german, Diploma thesis)

Date: 02.09.2009, 15:30-16:05
Location: Room 074 in the Fraunhofer IGD, Fraunhoferstrasse 5, S3|05
Speaker: Jordan Kavaldjiev (Advisor Tatiana von Landesberger)
Title: "Exploration und visuelle Suche in zeitabhängigen und strukurellen Finanzdaten - Konzept und prototypische Anwendung" (in german, Bachelor thesis)

Date: 25.08.2009, 16:30-17:00
Location: Room 073 in the Fraunhofer IGD, Fraunhoferstrasse 5, S3|05
Speaker: Sebastian Koch (Advisor Michael Goesele)
Title: "Visual Passwords"
Abstract: Traditional authentication systems are based on alphanumeric characters. More recently images have been used to allow users to authenticate themselves. Preceding systems used the same image during the procedure of creating a password and logging in. We present a system that can use different pictures showing the same object instead of using identical pictures. We further present the results of a user study showing the usability of the system. Participants used the system to create an own password and to login multiple times with their password.

GRIS Colloquium

Date: 21.07.2009, 16:30
Location: Room 073 in the Fraunhofer IGD Fraunhoferstrasse 5, S3|05
Speaker: Dr. Benjamin Bustos
Department of Computer Science (University of Chile)
Title: "On Nonmetric Similarity Search Problems in Complex Domains"
Abstract: For a long time, the database-oriented applications of similarity search employed the definition of similarity restricted to metric distances. Due to its topological properties, metric similarity can be effectively used to index a database that can be queried efficiently by so-called metric access methods. However, during the last years the demands for nonmetric similarity search have appeared, since the metric axioms became too restrictive for certain domain experts and their applications. In this talk, I will show examples of nonmetric similarity measures and their application domains. I will also present the existing database indexing techniques suitable for efficient nonmetric search.
Short bio: Benjamin Bustos is an Assistant Professor in the Department of Computer Science at the University of Chile. He is currently leading the Fondecyt (Chile) Project "Effective and Efficient Retrieval in Multimedia Databases" and is head of the PRISMA research group. His research interests are similarity search in multimedia databases, metric and multidimensional indexing, and multimedia information retrieval. He has a doctoral degree in natural sciences from the Universität Konstanz, Germany.

Date: 29.06.2009, 11:40
Location: Room 074 in the Fraunhofer IGD Fraunhoferstrasse 5, S3|05
Speaker: Dr. Robert Strzodka
Integrative Scientific Computing (MPI Saarbruecken)
Title: "Scientific Computing on Multi-GPU Systems"

Abstract: The advances in hardware functionality and programmability of graphics processors (GPUs) have greatly increased their appeal as add-on co-processors for scientific computing. However, for large-scale computations a desktop with a single GPU is not sufficient and parallel workstations or even clusters with multiple GPUs must be considered. This talk will address the challenges of these heterogeneous computing systems and in particular discuss the integration of hardware acceleration into parallel large scale software packages without requiring the user to change the existing application code.
Bio: Robert Strzodka is a senior researcher at the Max Planck Institut Informatik in Saarbrücken where he leads the independent research group Integrative Scientific Computing since 2007. The research focuses on efficient interactions of mathematic, algorithmic and architectural aspects in heterogeneous high performance computing. Previously, Robert was a visiting assistant professor in computer science at the Stanford University and until 2005 a postdoc at the Center of Advanced European Studies and Research in Bonn. He received his doctorate in numerical mathematics from the University of Duisburg-Essen in 2004.

Thesis presentations

Date: 24.6.2009, 10:00-10:30
Location: Room 074 in theFraunhofer IGD, Fraunhoferstrasse 5, S3|05
Speaker: Thorsten Franzel (Advisor Michael Goesele)
Title: "Verbesserte Structure-from-Motion Kalibrierung durch Bildverzerrung"

Date: 24.6.2009, 10:30-11:00
Location: Room 074 in the Fraunhofer IGD, Fraunhoferstrasse 5, S3|05
Speaker: Thomas Koch (Advisor Thomas Kalbe)
Title: "Interaktive Visualisierung variierender Isoflächen mit trivariaten Splines auf massiv parallelen Prozessoren"

Date: 09.6.2009, 16:30-17:00
Location: Room 073 in theFraunhofer IGD, Fraunhoferstrasse 5, S3|05
Speaker: Thorsten Franzel (Advisor Michael Goesele)
Title: "Verbesserte Structure-from-Motion Kalibrierung durch Bildverzerrung"

Date: 09.6.2009, 17:00-17:30
Location: Room 073 in the Fraunhofer IGD, Fraunhoferstrasse 5, S3|05
Speaker: Thomas Koch (Advisor Thomas Kalbe)
Title: "Interaktive Visualisierung variierender Isoflaechen mit trivariaten Splines auf massiv parallelen Prozessoren"

Date: 05.6.2009, 11:00-11:30
Location: Raum 220 in the Fraunhofer IGD, Fraunhoferstrasse 5, S3|05
Speaker: Wolfgang Kleine (Betreuerin Svenja Kahn)
Title: "Extraktion eines 3D-Kantenmodells aus Tiefenbildern"

GRIS Colloquium

Date: 04.06.2009, 16:30
Location: Room 074 in the Fraunhofer IGD, Fraunhoferstrasse 5, S3|05
Speaker: Prof. Dr.-Ing. Holger Theisel
                  Visual Computing Group (University of Magdeburg)
Title: "Streak Surfaces for Flow Visualization"


Technische Universität Darmstadt

Interactive Graphics Systems Group

Fraunhoferstr. 5
64283 Darmstadt

Tel. +49 6151 155 679

icon email office@gris.tu-

A A A | Print Print | Legal note Legal Notes | Contact Contact
to topto top