Seminars/Colloquia/Invited Talks

Seminars

Eran Borenstein

Face Detection - A Compositional Approach

PLACE:	Clark 314
EVENT:	CIS Seminar Series
DATE:	November 14, 2006
TIME:	1:00 - 2:00 PM

Abstract

We are attempting to address the "ROC gap" between human and machine performance in face detection by constructing probabilistic generative models and approaching the recognition problem from the Bayesian point of view. I will describe an approach based on hierarchy and reusability that we call compositional. Two aspects of the compositional approach that differ from most other Bayesian approaches are (1) it is non-Markovian (unlike a "Bayes net," or a probabilistic context-free grammar), and (2) it is fully generative in that the observation model (distribution on the data) is a distribution on grey-level images (pixels) rather than on a priori features. The non-Markovian prior is formulated as a distribution on parses (a.k.a. interpretations), where a parse identifies faces and pieces of faces in unconstrained images. The observation model then bridges the data to the interpretation. I will present, briefly, our definition of a parse and our compositional distribution on parses, and then present an in-depth description of the observation model. Issues addressed include learning, normalization, and sampling. I will conclude by showing sample images of faces from the full generative model.
This is a joint work with Stuart Geman, Elie Bienenstock and Wei Zhang.

Brief biography

Eran Borenstein received his B.Sc degree in electrical engineering from the Technion - Israel Institute of Technology in 1998, and completed his M.Sc and PhD from the Department of Computer Science and Applied Mathematics at the Weizmann Institute of Science in 2004.
Borenstein pursued a parallel postdoctoral fellowship at the Mathematical Sciences Research Institute (MSRI) and the Electrical Engineering and Computer Science Department, both of University of California at Berkeley during 2004-2005. He is now a postdoctoral fellow at Brown University's Department of Applied Mathematics.
His research focuses on computer vision and the computational modeling of human vision. His work addresses the intimately related problems of image segmentation, grouping and recognition.