The Weizmann Institute of Science
Faculty of Mathematics and Computer Science
Vision and Robotics Seminar
Lihi Zelnik-Manor
Department of Electrical Engineering
Caltech
will speak on
Decomposition of video data
Abstract:
Dynamic scenes captured by a moving camera generate rich and complex dynamic data. To understand the content of the data
and facilitate its processing one should decompose it into smaller meaningful chunks.
I will start by presenting an approach to modeling non-rigid changes in a video sequence. I will then show that to obtain
decomposition of such data (as well as in many other applications) one needs a clustering tool which can operate on affinity
relations which are beyond dyadic (pairwise) relations - e.g., triadic, tetradic, or relations of higher orders. I will propose
a two-step approach for solving this hyper partitioning problem. First, I will show how a hypergraph can be approximated by a
weighted graph and then apply a graph-partitioning algorithm to obtain the final decomposition. Accompanying theoretical analysis
proves the superiority of the suggested approximation algorithm over the existing approaches. I will further suggest a
graph-partitioning algorithm which tunes itself to the appropriate scale of the data, handles multiple scales, is robust to
irregular background clutter and automatically finds the number of groups.
Throughout the talk, I will show the applicability of the suggested decomposition approaches to a wide variety of problems
in video analysis, including temporal factorization, image segmentation, and identity recognition under varying illumination.
The lecture will take place in the
Lecture Hall, Room 1, Ziskind Building
on Sunday, July 10, 2005
15:00 - 16:00