A learned overcomplete sparseness and IGMRF based regularization framework for dense disparity estimation

Abstract

In this work, we propose to use an Inhomogeneous Gaussian Markov Random Field (IGMRF) and sparsity based priors in a regularization framework in order to estimate the dense disparity map. The IGMRF prior captures the spatial variation among disparities locally as well as it preserves sharp discontinuities. The sparsity prior captures the additional structure such as sparseness in the disparity map. The sparseness of the disparities are represented over the overcomplete dictionary which is learned from the estimated disparity map of the given stereo pair, using K-singular value decomposition (K-SVD) algorithm. The dictionary atoms are adaptive to the disparities of the given stereo pair. The sparse representation of disparities is used as a prior which is combined with the IGMRF prior in an energy minimization framework for estimating the disparity map. Disparity map is estimated using a two phase, iterative algorithm. In phase one, IGMRF parameters are computed at each pixel location and the dictionary is learned as well as the sparseness of disparities are obtained while keeping the disparity map fixed, and in phase two, disparity map is estimated by keeping the other parameters fixed. Experimental results on the standard dataset demonstrate the effectiveness of the proposed approach.