A Markov Random Field Model and Method to Image Matching

In this paper, the correspondence problem is solved by minimizing an energy functional using a stochastic approach. Our procedure generally follows Geman and Geman's Gibbs sampler for Markov Random Fields (MRF). We propose a transition generator to generate and explore states. The generator allows constraints such as epipolar, uniqueness, and order to be imposed. We also propose to embed occlusions in the model. The energy functional is designed to take into account resemblance, continuity, and number of occlusions. The disparity and occlusion maps as modeled by their energy functional, i.e., as a Gibbs-Boltzmann distribution, are viewed as a MRF where the matching solution is an optimal state.