home · mobile · calendar · colloquia · 2000-2001 · 

Colloquium - Mulligan

Stereo Vision for Telepresence: Methods and Metrics
GRASP Laboratory, University of Pennsylvania

Advances in network speed and bandwidth have created opportunities for previously unrealizable Computer Vision and Graphics applications such as Networked Augmented Reality and Telepresence. Making these applications a reality requires real-time environment sensing to acquire compelling scene models for rendering and user interaction at remote locations. Computer Vision techniques have recently been embraced by the Graphics community as a means of acquiring such large-scale, realistic world models. For online multi-media applications such as Tele-Immersion, we are concerned with the `immersive feeling' of the rendered models, including the quality of depth points in terms of accuracy and absence of outliers, and the responsiveness of the system necessary for communication and interaction among users. Challenges for a real-time stereo system which must deliver this level of environment scanning include trade-offs for quality versus disparity search, and prediction and modeling of an unstructured environment. Further, for systems requiring high network bandwidth, representations must take into account application based network exploitation and quality of service.

I will describe a stereo vision system I developed for a working prototype tele-immersion system including novel algorithms, and an extensive dataset of calibrated trinocular stereo images and corresponding laser range scans (which act as ground truth) being used to evaluate stereo reconstructions for immersive environments.

Hosted by Clayton Lewis.

Department of Computer Science
University of Colorado Boulder
Boulder, CO 80309-0430 USA
May 5, 2012 (14:13)