Speaker:  Chetan Arora, IIIT Delhi

Date/Time/Venue: Feb 20, 12noon, SIT 001

Title: Inference Algorithms for MRF-MAP Problems in Computer Vision with Extremely Large Cliques

Abstract:  Several tasks in computer vision and machine learning can be modelled as MRF-MAP inference problems. Using higher order potentials to model complex dependencies can significantly improve the performance. While the general MRF-MAP optimization problem is computationally hard, algorithms have been developed which have acceptable computational complexity when the potentials are submodular. We have shown that this optimization problem can be solved by network flow techniques, which successfully pushed the tractability of state of the art from cliques of size 4 to 16 for problems containing millions of cliques. But, the solution is practical only when potentials are defined over small cliques. When the cliques are of large size we show a block co-ordinate descent based methodology which generalizes the standard Min Norm Point algorithm to handle very large cliques of size even upto 1000. We also show a hybrid approach which combines our flow based algorithm with the generalized Min Norm Point algorithm to handle problems containing millions of small, and large cliques.

Speaker's Bio:  Chetan Arora received his B.Tech in Electrical Engineering and Ph.D. in Computer Science, both from IIT Delhi in 1999 and 2012 respectively. He did his Post Doctoral Research with Prof. Shmuel Peleg from 2012-2014 in Hebrew University, Israel. Since 2014, he is an Assistant Professor in CS at IIIT Delhi. Chetan has spent over 10 years in industry, where he cofounded 3 startups, along with his counterparts in Japan and Israel, all working on computer vision products coming out of latest research ideas.