Liste der Publikationen
A Fast, Massively Parallel Solver for Large, Irregular Pairwise Markov Random Fields
High Performance Graphics 2016
High-Performance Graphics (HPG) <8, 2016, Dublin, Ireland>
Given the increasing availability of high-resolution input data, today's computer vision problems tend to grow beyond what has been considered tractable in the past. This is especially true for Markov Random Fields (MRFs), which have expanded beyond millions of variables with thousands of labels. Such MRFs pose new challenges for inference, requiring massively parallel solvers that can cope with large-scale problems and support general, irregular input graphs. We propose a block coordinate descent based solver for large MRFs designed to exploit many-core hardware such as recent GPUs. We identify tree-shaped subgraphs as a block coordinate scheme for irregular topologies and optimize them efficiently using dynamic programming. The resulting solver supports arbitrary MRF topologies efficiently and can handle arbitrary, dense or sparse label sets as well as label cost functions. Together with two additional heuristics for further acceleration, our solver performs favorably even compared to modern specialized solvers in terms of speed and solution quality, especially when solving very large MRFs.
Constructive Roofs from Solid Building Primitives
Transactions on Computational Science XXVI
International Conference on Cyberworlds (CW) <13, 2014, Santander, Spain>
The creation of building models has high importance, due to the demand for detailed buildings in virtual worlds, games, movies and geo information systems. Due to the high complexity of such models, especially in the urban context, their creation is often very demanding in resources. Procedural methods have been introduced to lessen these costs, and allow to specify a building (or a class of buildings) by a higher level approach, and leave the geometry generation to the system. While these systems allow to specify buildings in immense detail, roofs still pose a problem. Fully automatic roof generation algorithms might not yield desired results (especially for reconstruction purposes), and complete manual specification can get very tedious due to complex geometric configurations. We present a new method for an abstract building specification, that allows to specify complex buildings from simpler parts with an emphasis on assisting the blending of roofs.
Joint Optical Flow and Temporally Consistent Semantic Segmentation
Computer Vision - ECCV 2016 Workshops
European Conference on Computer Vision (ECCV) <14, 2016, Amsterdam, The Netherlands>
The importance and demands of visual scene understanding have been steadily increasing along with the active development of autonomous systems. Consequently, there has been a large amount of research dedicated to semantic segmentation and dense motion estimation. In this paper, we propose a method for jointly estimating optical flow and temporally consistent semantic segmentation, which closely connects these two problem domains and leverages each other. Semantic segmentation provides information on plausible physical motion to its associated pixels, and accurate pixel-level temporal correspondences enhance the accuracy of semantic segmentation in the temporal domain. We demonstrate the benefits of our approach on the KITTI benchmark, where we observe performance gains for flow and segmentation. We achieve state-of-the-art optical flow results, and outperform all published algorithms by a large margin on challenging, but crucial dynamic objects.
new/s/leak - Information Extraction and Visualization for Investigative Data Journalists
The 54th Annual Meeting of the Association for Computational Linguistics
Annual Meeting of the Association for Computational Linguistics <54, 2016, Berlin, Deutschland>
We present new/s/leak, a novel tool developed for and with the help of journalists, which enables the automatic analysis and discovery of newsworthy stories from large textual datasets. We rely on different NLP preprocessing steps such named entity tagging, extraction of time expressions, entity networks, relations and metadata. The system features an intuitive web-based user interface based on network visualization combined with data exploring methods and various search and faceting mechanisms. We report the current state of the software and exemplify it with the WikiLeaks PlusD (Cablegate) data.
Rapid, Detail-Preserving Image Downscaling
ACM Transactions on Graphics
Conference and Exhibition on Computer Graphics and Interactive Techniques in Asia (SIGGRAPH ASIA) <9, 2016, Macao>
Image downscaling is arguably the most frequently used image processing tool. We present an algorithm based on convolutional filters where input pixels contribute more to the output image the more their color deviates from their local neighborhood, which preserves visually important details. In a user study we verify that users prefer our results over related work. Our efficient GPU implementation works in real-time when downscaling images from 24M to 70 k pixels. Further, we demonstrate empirically that our method can be successfully applied to videos.
SeismoTracker: Upgrade Any Smart Wearable to Enable a Sensing of Heart Rate, Respiration Rate, and Microvibrations
Proceedings of the 2016 CHI Conference Extended Abstracts on Human Factors in Computing
Conference on Human Factors in Computing Systems (CHI) <34, 2016, San Jose, CA, USA>
In this paper we present a method to enable any smart Wearable to sense vital data in resting states. These resting states (e.g. sleeping, sitting calmly, etc.) imply the presence of low-amplitude body-motions. Our approach relies on seismocardiography (SCG), which only requires a built-in accelerometer. Compared to commonly applied technologies, such as photoplethysmography (PPG), our approach is not only tracking heart rate (HR), but also respiration rate (RR), and microvibrations (MV) of the muscles, while being also computational inexpensive. In addition, we can calculate several other parameters, such as HR variability and RR variability. Our extracted vital parameters match with the vital data gathered from clinical state-of-the art technology. These data allow us to gain an impression on the user's activity, quality of sleep, arousal and stress level over the whole day, week, month, or year. Moreover, we can detect whether a device is actually worn or doffed, which is crucial when connecting such data with health services. We implemented our method on two current smartwatches: a Simvalley AW420 RX as well as on a LG G Watch R and recorded user data for several months. A web platform enables to keep track of one's data.
Shading-Aware Multi-View Stereo
Computer Vision - ECCV 2016. Proceedings Part III
European Conference on Computer Vision (ECCV) <14, 2016, Amsterdam, The Netherlands>
We present a novel multi-view reconstruction approach that effectively combines stereo and shape-from-shading energies into a single optimization scheme. Our method uses image gradients to transition between stereo-matching (which is more accurate at large gradients) and Lambertian shape-from-shading (which is more robust in flat regions). In addition, we show that our formulation is invariant to spatially varying albedo without explicitly modeling it. We show that the resulting energy function can be optimized efficiently using a smooth surface representation based on bicubic patches, and demonstrate that this algorithm outperforms both previous multi-view stereo algorithms and shading based refinement approaches on a number of datasets.
Supporting Collaborative Political Decision Making - An Interactive Policy Process Visualization System
VINCI 2016. The 9th International Symposium on Visual Information Communication and Interaction
International Symposium on Visual Information Communication and Interaction (VINCI 2016) <9, 2016, Dallas, Texas>
The process of political decision making is often complex and tedious. The policy process consists of multiple steps, most of them are highly iterative. In addition, different stakeholder groups are involved in political decision making and contribute to the process. A series of textual documents accompanies the process. Examples are official documents, discussions, scientific reports, external reviews, newspaper articles, or economic white papers. Experts from the politi- cal domain report that this plethora of textual documents often exceeds their ability to keep track of the entire policy process. We present PolicyLine, a visualization system that supports different stakeholder groups in overview-and-detail tasks for large sets of textual documents in the political decision making process. In a longitudinal design study conducted together with domain experts in political decision making, we identfied missing analytical functionality on the basis of a problem and domain characterization. In an iterative design phase, we created PolicyLine in close collaboration with the domain experts. Finally, we present the results of three evaluation rounds, and reect on our collaborative visualization system.
The Cityscapes Dataset for Semantic Urban Scene Understanding
29th IEEE Conference on Computer Vision and Pattern Recognition. Proceedings
IEEE Conference on Computer Vision and Pattern Recognition (CVPR) <2016, Las Vegas, Nevada, USA>
Visual understanding of complex urban street scenes is an enabling factor for a wide range of applications. Object detection has benefited enormously from large-scale datasets, especially in the context of deep learning. For semantic urban scene understanding, however, no current dataset adequately captures the complexity of real-world urban scenes. To address this, we introduce Cityscapes, a benchmark suite and large-scale dataset to train and test approaches for pixel-level and instance-level semantic labeling. Cityscapes is comprised of a large, diverse set of stereo video sequences recorded in streets from 50 different cities. 5000 of these images have high quality pixel-level annotations; 20 000 additional images have coarse annotations to enable methods that leverage large volumes of weakly-labeled data. Crucially, our effort exceeds previous attempts in terms of dataset size, annotation richness, scene variability, and complexity. Our accompanying empirical study provides an in-depth analysis of the dataset characteristics, as well as a performance evaluation of several state-of-the-art approaches based on our benchmark.
Visual-Interactive Search for Soccer Trajectories to Identify Interesting Game Situations
Visualization and Data Analysis 2016
Visualization and Data Analysis (VDA) <2016, San Francisco, CA, USA>
Electronic Imaging, 1
Recently, sports analytics has turned into an important research area of visual analytics and may provide interesting findings, such as the best player of the season, for various kinds of sports. Soccer is a very popular and tactical game, which also attracted great attention in the last few years. However, the search for complex game movements is a very crucial and challenging task. We present a system for searching trajectory data in soccer matches by means of an interactive search interface that enables the user to sketch a situation of interest. Furthermore, we apply a domain specific prefiltering process to extract a set of local movement segments, which are similar to a given sketch. Our approach comprises single-trajectory, multi-trajectory, and event-specific search functions based on two different similarity measures. To demonstrate the usefulness of our approach, we define a domain specific task analysis and conduct a case study together with a domain expert from FC Bayern München by investigating a real-world soccer match. Finally, we show that multi-trajectory search in combination with event-specific filtering is needed to describe and retrieve complex moves in soccer matches.