Stereo matching by using self-distributed segmentation and massively parallel GPU computing

Abstract

As an extension of using image segmentation to do stereo matching, firstly, by using self-organizing map (som) and K-means algorithms, this paper provides a self-distributed segmentation method that allocates segments according to image’s texture changement where in most cases depth discontinuities appear. Then, for stereo, under the fact that the segmentation of left image is not exactly same with the segmentation of right image, we provide a matching strategy that matches segments of left image to pixels of right image as well as taking advantage of border information from these segments. Also, to help detect occluded regions, an improved aggregation cost that considers neighbor valid segments and their matching characteristics is provided. For post processing, a gradient border based median filter that considers the closest adjacent valid disparity values instead of all pixels’ disparity values within a rectangle window is provided. As we focus on real-time execution, these time-consumming works for segmentation and stereo matching are executed on a massively parallel cellular matrix GPU computing model. Finaly, we provide our visual dense disparity maps before post processing and final evaluation of sparse results after post-processing to allow comparison with several ranking methods top listed on Middlebury.