December 7, 2010 - Richard Hartley

Title: Provably Optimal Algorithms in Computer Vision

 

Abstract:
The usual algorithms in Computer Vision Structure and Motion are large non convex problems, often involving large numbers of variables (in excess of a million). Through a variety of simple techniques, it is possible to find initial solutions that will serve for initialization of iterative algorithms with good results.  Nevertheless, it is interesting to find algorithms that are provably optimum, finding a guaranteed global minimum. This talk gives a summary of work in this area, involving a variety of techniques, including L-infinity optimization, branch and bound and convex verification techniques.