"Planarity Matters: Variational inference in Planar Markov Random Fields and Variational Approach to Planar Correlation Clustering"

Julian Yarkony, PhD Student ICS, UC, Irvine

February 23rd (Thursday), 11:00am
Harold Frank Hall, Rm 4164 (ECE Conference Room)

Variational inference in Planar Markov Random Fields: MAP Inference in Markov random fields (MRFs) is of great importance for many vision applications. MRFs which appear in computer vision tend to be planar in structure. We demonstrate various means to take advantage of reductions of MAP estimation to perfect matching in a dual decomposition framework. We show how this can be applied to binary MRFs and non-binary MRFs when planarity is present. Our algorithms provide upper and lower bounds on the MAP which can be tightened iteratively and can be interpreted with regard to the marginal polytope.

Variational Approach to Planar Correlation Clustering: Grouping pixels in an image into regions is a key problem in computer vision research. We frame the problem as a correlation clustering problem on a planar graph. In this talk a new algorithm for computing the ground state (MAP) for planar correlation clustering problems is introduced. Our algorithm substitutes an intractable minimization for a tractable maximization of a variational lower bound then maps the lower bound to a solution. In practice our lower bound and the energy of the final solution have the same value providing a guarantee of global optimality. The key advantage to our work is that the algorithm will produce clustering of the super-pixels in an image without receiving the number of clusters as input.

About Julian Yarkony:

Julian Yarkony is a 5th year PhD student under Professors Fowlkes and Ihler at the University of California at Irvine. His research focuses on the development of approximation algorithms for applications in image analysis. He is currently focusing on hierarchical clustering in images under a variational framework. His published results concern the creation of effective MAP estimators for Markov random fields especially planar Markov random fields. In his spare time he enjoys fine arts, and the ocean.

Hosted by: Professor B.S.Manjunath