MAP-Based Stochastic Diffusion for Stereo Matching and Line Fields Estimation

摘要

This paper proposes a stochastic approach to estimate the disparity field combined with line field. In the maximum a posteriori (MAP) method based on Markov random field (MRF) model, it is important to optimize and converge the Gibbs potential function corresponding to the perturbed disparity field. The proposed optimization method, stochastic diffusion, takes advantage of the probabilistic distribution of the neighborhood fields to diffuse the Gibbs potential space iteratively. By using the neighborhood distribution in the non-random and non-deterministic diffusion, both the estimation accuracy and the convergence speed are improved. In the paper, the hierarchical stochastic diffusion is also applied to the disparity field. The hierarchical approach reduces the memory and computational load, and increases the convergence speed of the potential space. The paper also proposes an effective configuration of the neighborhood to be suitable for the hierarchical disparity structure. According to the experiments, the stochastic diffusion shows good estimation performance. The line field improves the estimation at the object boundary, and coincides with the object boundary with the useful contours. The stochastic diffusion is applicable to any kind of field estimation given the appropriate definition of the field and MRF models.