Challenges in 3D Scene Acquisition and Reconstruction
Organizers: Hendrik Lensch, Christian Theobalt, Michael Wand
Abstract:
A realistic model of a real-world scene should include a representation of the scene's appearance which is influenced by its geometry, reflection properties, and its motion. The reconstruction of each individual of these properties from sensor data is by itself a demanding task. In this workshop we will address state-of-the-art acquisition and reconstruction techniques and identify their limitations. In particular we want to discuss how techniques from statistical data analysis and 3D sensor technology, such as camera and projector arrays, can be advanced and possibly integrated into a joint framework. By means of such an integration, novel problems could be attacked such as spatio-temporally consistent geometry acquisition of dynamic scenes, joint geometry and appearance acquisition, or space-time stereo with multi-camera multi-projector arrays. The workshop will start with short presentations by the organizers. Thereafter, there will be a discussion on open challenges an possible project ideas.
Real-time computer vision
Organizers: Robert Strzodka and Bodo Rosenhahn
Abstract:
Computer vision lies at the intersection of multiple disciplines and draws on the power of psychological insights, artificial intelligence, mathematical modeling, algorithmic design, image processing and system integration to attack the very difficult problem of visual understanding. While the quality and flexibility of the human visual system is out of reach, this problem is extremely hard for camera equiped computer systems or robots, even for restricted environments and tasks. Many components have to work together towards a common goal, and often promising techniques are infeasible because of their enormous computing requirements, especially when the analysis must take place in real-time.
Although the involved algorithms operate on data dependent regions, there is still a lot of parallelism present which can be exploited on high performance parallel processors like Cell, GPU or FPGA. Individual components of the processing chain have thus been accelerated significantly on such architectures but the challenge lies in the integration of these hardware components into a complex framework and their smooth interaction.
The workshop will discuss possibilities for parallelization of different computer vision tasks such as feature extraction and tracking, segmentation, registration or motion computation. Introductory presentations will offer an overview of these computer vision problems and explain the potential of graphics processors to accelerate certain tasks. In the following discussion we want to identify the strategies to integrate the promising hardware solvers into such vision systems to achieve real-time performance.
Organizers: Joachim Giesen and Stefan Funke
Abstract:
Simulation of molecule interactions via force field calculations allows to get fundamental insight into the dynamic behavior and function of molecules. Typical simulation systems consider an extremely fine discretization of time (in the range of femto-seconds) and step by step calculate for each atom the external forces experienced. Based on this force field computation, the movement of the atom and the whole molecule as such is determined. An individual step of this computation is rather costly, even if the naive approach of considering all pairwise interactions between the atoms is accelerated by methods like multipole expansion. Our goal is to structure the results produced by such simulations. The expected benefit is two-fold: on one hand, insights into the structure of molecule dynamics allows for more efficient simulation methods if the same---or similar---molecules are considered in another simulation. On the other hand, the sheer amount of data produced by current simulation systems could easily be reduced via adaptive compression schemes if the atoms' movements were better understood. The latter is an important issue since currently many of the simulation results are thrown away just because for the lack of permanent storage. Relevant tools and techniques to work on this problem come from the areas of computational geometry and external memory algorithms.
Multiview Video Streaming via Content Coding-and-Delivery Networks
Organizers: Marcus Flierl and Pierpaolo Baccichet
Abstract:
Multiview video contents on the Internet will grow with the number of sports events that indulge their audience with live video from multiple view-points of the arena. The delivery of multiple live videos from a large number of cameras to many remote clients will be a challenge for the bandwidth constrained Internet. Obviously, to reduce the bandwidth, the temporal correlation of each video signal can be exploited at each camera. But for highly correlated multiview video, the bandwidth can be reduced further by exploiting the correlation across the camera views. To accomplish the latter, we will consider a Content Coding-and-Delivery Network (CCDN). The CCDN will consist of both multiple video source nodes that provide the compressed camera signals as well as proxies that exploit the view correlation and balance the load of heterogeneous requests from the clients. In contrast to the bandwidth-constrained Internet, the CCDN is a high-speed network that interconnects source nodes and proxies.
The workshop will evaluate the challenges arising from the deployment of a CCDN. In particular, the representation of the video signals as well as the construction of the delivery network will be addressed. The construction will be constrained by the processing capabilities of the proxies, the available network bandwidth, and the set of requests from the clients. The problem will be particularly challenging if we permit client requests that range from a single view to multiple views for 3D displays.