Augmenting graph cut with TV-L1 approach for robust stereo matching

Stereo matching plays a significant role in various vision based systems including 3D reconstruction, robot localization, mapping and navigation and it has been an intense area of research for many years. In the last decade, with the increase in processing power and reduction in the camera cost, the stereo systems have found many new applications. In this paper, we present a stereo matching algorithm which computes the disparities with sub-pixel accuracy. The proposed method combines graph cut based stereo matching method with total variation approach to achieve improved accuracy. We demonstrate the robust performance of the algorithm on the large scale Middlebury benchmark. The algorithm has also been tested on real data and the results compare favorably with the similar category state-of-the-art algorithms.