Adaptive support weight aggregation in segmentationsfor stereo matching

Computing the dissimilarity between the central pixel and other pixels within the support window and determining the size and shape of the support window are core ideas of the adaptive support weight approach in local methods of stereo matching. In this paper we focus on the latter issue and propose a new strategy to find that window. First, the reference image is segmented into a set of irregular regions; second, for each pixel, the support weight and aggregation is calculated within the segmentation it belongs to; third, all disparities are selected by winner take all strategy. Through the segmentation technique, the proposed approach implements an adaptive support window without the size and shape limitation. Experiment results based on the Middlebury benchmark show that, compared with the original approach, the new approach effectively improves the accuracy of result disparity maps.