Deep Belief Propagation for Higher-Order MRF Inference

In this work, we develop an efficient belief propagation algorithm, deep belief propagation, for the inference of a class of higher-order MRFs. Recent success in using MRF to solve computer vision problems has also revealed its deficiency in capturing long-range variable interactions. This has motivated people to study the cliques with large number of variables, i.e., higher-order cliques. Another motivation for studying higher-order cliques comes from the long-range sensing, where sensed signals record the interaction between a large number of random variables. These higher-order MRFs have posed challenges to the corresponding inference problem. In this work, it is shown that although these higher-order cliques cannot be further factorized as smaller cliques, there exists another kind of "deep" factorization structure of these cliques, which can be explored by dynamic programming. The developed dynamic programming algorithm can be further implemented using message passing, resulting in the deep belief propagation algorithm for the higher-order MRF inference.

Publications:

• In preparation.