Login
Computer Science @ Rochester
Monday, March 26, 2001
11:00 AM
CSB 209
Jim Rehg
Compaq CRL
Motion Capture from Movies
Video is the primary archival source for human movement, with examples ranging from sports coverage of Olympic events to dance routines in Hollywood movies. If the human figure could be tracked reliably in unconstrained monocular video, much of this archive could be unlocked for analysis. Significant technical challenges exist, however, due to the complexity of human movement, the variability of human appearance, and the loss of 3-D information.

My talk will describe some recent progress in modeling and estimating figure motion from monocular video. Two important themes are the separation of 2-D (registration) and 3-D (reconstruction) effects in kinematic modeling, and the role of learning in dynamic modeling. In particular, I will describe a framework for learning switching linear dynamic system models from data and show its application to figure motion. I will describe some applications to video editing and computer animation currently underway at Compaq Research.