ColloquiaTalk by Olga Veksler on Wed. April 17

Margery Ishmael marge at cs.uchicago.edu
Tue Apr 9 13:27:14 CDT 2002


Wednesday, April 17
at 2:30 pm
in Ryerson 251

"Graph Algorithms for Stereo Correspondence"
OLGA VEKSLER
NEC Research Institute
http://external.nj.nec.com/homepages/olga/

Graph algorithms have recently gained popularity in computer vision. They 
provide powerful combinatorial optimization tools for many vision 
problems.  In the last four years I have been developing graph based 
algorithms for motion, segmentation, edge detection, image restoration, and 
stereo correspondence.  In this talk I will focus on stereo, for three 
reasons.  First, stereo shares a common structure with many other vision 
problems, and presents a similar set of challenges.  Second, graph 
algorithms have achieved significant recent progress in solving 
stereo.  Last, an objective evaluation of stereo algorithms on real images 
is now possible using the Middlebury stereo database of dense ground truth.

Traditional approaches to stereo correspondence can be roughly divided in 
two groups: local and global.  A central problem for the local methods is 
selecting an appropriate window shape.  I will show how the minimum ratio 
cycle algorithm can optimize the search for window shapes.  A central 
problem in the global approach is the optimization of the objective 
function.  I will show how graph cut algorithms can efficiently compute 
provably good approximations to a certain class of objective 
functions.  The local approach is faster but works well only for high 
texture images.  The global approach is slower but it can handle low 
texture images.

Host: Prof. Partha Niyogi

*The talk will be followed by refreshments in Ryerson 255*
Persons who need assistance should call 773.834.8977

=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=
Margery Ishmael
Secretary to the Chairman, Department of Computer Science
The University of Chicago
tel. 773.834.8977  fax. 773.702.8487
=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=





More information about the Colloquium mailing list