|Stochastic Systems Group|
This talk explores meta descriptions for video that focus on objects and their motions rather than on pixels and frames. These meta constructs provide very efficient video representations with large data reduction gains. After a brief overview of some of our results with 2D representations, in particular,with videos with human walkers, I will focus on recovering the 3D structure (3D shape and 3D motion) of objects from monocular video sequences. I present the surface based factorization method that solves this inverse problem. I describe a rank 1 matrix factorization algorithm that exploits subspace constraints and leads to an efficient solution to the 3D structure recovery problem. I will present results with real videos that illustrate the performance of the surface based factorization approach.
Problems with this site should be emailed to firstname.lastname@example.org