A learning based approach for dense stereo matching with IGMRF prior

In this paper, we propose a learning based approach for solving the problem of dense stereo matching problem using edge preserving regularization prior. Given the test stereo pair and a training database consisting of disparity maps estimated using multiple views stereo images and their corresponding ground truths, we obtain the disparity map for the test set. We first obtain an initial disparity estimate by learning the disparities from the available database. A new learning based approach is proposed for obtaining the initial estimate that uses the estimated and the true disparities. Since the disparity estimation is an ill posed problem, we obtain the final disparity map using a regularization framework. The prior model for the disparity map is chosen as an Inhomogeneous Gaussian Markov Random Field (IGMRF). Assuming that the spatial variations among the disparity values captured in an initial estimate correspond to the variations in true disparities, we obtain the IGMRF parameters at every pixel location using the initial estimate. A graph cuts based method is used to optimize the energy function in order to obtain the global minimum. Experimental results on the standard dataset demonstrate the effectiveness of the proposed approach.