EAGER: Solving Markov Random Fields with Mutual Exclusion Constraints

Temple University, Philadelphia PA

Investigators

Abstract

This project explores a new way to increase the expressive power of Markov Random Fields (MRF) while at the same time improving the computing efficiency and the quality of solutions. The research focuses on utilizing quadratic mutual exclusion constraints (QMCs) expressed in quadratic equality form. Current approaches to increasing the expressive power of MRFs face a very challenging problem of higher computing time. In contrast, this approach is able to restrict the solution search space with QMCs, which in turn not only leads to significantly better solutions but also to reduced computing time. Many problems in computer vision, including but not limited to image and video segmentation, stereo, and image restoration, object detection and recognition, tracking, and activity recognition, are formulated as optimization problems involving inference of the maximum a posteriori (MAP) solution of a Markov Random Field (MRF). QMCs are more general than mutex constraints expressed in a linear equality form. Hence QMCs offer increased expressive power to more accurately model many computer vision problems. This property is particularly important when unary and binary MRF potentials are unreliable and uninformative, which is the rule rather than an exception in real applications. Hence, this project can increase the ability of computer vision systems to broaden their application scope, ranging from image retrieval to computer vision systems on mobile robots.