*************************************************************************

                Department of Mathematical Sciences                       
                   The Johns Hopkins University                           
                                                                        
                              SEMINAR                                
                                                                        
*************************************************************************

Professor Donald Geman                            February 10, 2000
Department of Mathematics and Statistics          304 Whitehead Hall
University of Massachusetts                       NO PRESEMINAR
                                                  Refreshments: 3:30 p.m.
                                                  Seminar: 4:00 p.m.

*************************************************************************

                 COARSE-TO-FINE COMPUTATIONAL VISION

*************************************************************************

                             ABSTRACT

I will summarize a research program in computational vision, originally
motivated by the startling efficiency of adaptive testing in parlor games
such as "twenty questions."  A natural framework for problems such as
visual tracking and object detection is then statistical and
information-theoretic, with performance measured by the amount of
computation necessary to reach a given level of accuracy.  This leads
naturally to the selection of image functionals ("features") based solely
on relative likelihoods, having no a priori semantic or geometric
interpretation, and to very sparse, discrete object representations.  It
also leads to highly coarse-to-fine processing, in both the complexity of
the features and in the exploration of poses and resolutions.  Moreover,
since image regions are explored according to their information content,
the spatial distribution of computation is very skewed.  After briefly
describing some model-based work using entropy coding for road-tracking and
image retrieval, I will discuss ongoing, learning-based experiments on
measuring lesions in MRI brain scans and on detecting instances from a
generic object class (e.g., a face) in natural scenes.  Two key
mathematical issues are "decomposable events" and efficient testing
strategies.

*************************************************************************