Abstract: We propose a local stereo mathing algorithm, which relies on combination initial matching via distinctive dissimilarity measure with post-processing by greedy disparity estimation. Based on a dissimilarity measure derived from the stereo itself, the matching costs are first constructed to obtain initial disparities for both views. Then, in order to address complex ambiguity in large occluded regions, a novel disparity estimation procedure including unreliable disparity detection, greedy disparity filling and multi-directional weighted least-square-error fitting is presented to improve matching performance. Experimental results indicate that the proposed method can reduce false matching in ambiguous regions to generate a dense disparity accurately and efficiently.

Key words: stereo matching, ambiguity, dissimilarity measure, greedy algorithm, large occlusion