Seminars/Colloquia/Invited Talks

Seminars

Iasonas Kokkinos

Towards Bridging Bottom-Up and Top-Down Vision with Hierarchical, Compositional Models

PLACE:	Clark 314
EVENT:	CIS Seminar
DATE:	February 26, 2008
TIME:	1:00 - 2:00 PM

Abstract

One of the most interesting questions in computer vision is how one can combine the e±ciency of bottom-up computation with the tractability of top-down inference. In this talk I will present recent work on the intimately related problem of 'object parsing', namely recursively composing the structures constituting an object, starting from elementary image information. The main contribution is controlling the problem's combinatorial complexity by focusing on structures computed from the bottom-up based on whether they may lead to a parse of the whole object. This is accomplished in the principled setting of the A* algorithm, by using top-down information as guidance to prioritize search. Specifically, I introduce a formulation of composition rules that allows to rapidly compute coarse solutions for hierarchical compositional models. This is used by A* as a lower bound of the parsing cost in order to rule out unpromising search directions. This approach is experimentally validated using natural images containing heavy clutter, and is shown to deliver accurate object parses on a challenging object detection benchmark.

Brief Biography:

Iasonas Kokkinos obtained his Diploma and Ph.D. degrees from the School of Electrical and Computer Engineering at the National Technical University of Athens in 2001 and 2006 respectively and is currently a postdoctoral scholar at UCLA. His research combines mathematics, computer science and statistics, and has been on nonlinear dynamical models of speech, biologically motivated models for vision, texture segmentation and object recognition. His current research focus is on jointly solving low- and high- level computer vision problems in a probabilistic framework. He has explored aspects of this problem including object-based segmentation, combining bottom-up and top-down computation for object detection, learning deformation models and object parsing.