Local smoothness enforced cost aggregation on weighted spanning trees

Cost aggregation is one of the key steps in stereo matching process. To improve the accuracy especially in the challenging regions such as boundaries and occluded regions, this paper presents a new cost aggregation approach on minimum spanning trees called robust weighted minimum spanning tree (W-MST). Our method applies not only color features but also gradient information to compute weight in order to improve the matching accuracy in textureless regions. Furthermore, we combine local smoothness with cost aggregation to address the shortage of the “fronto-parallel plane” and obtain accurate disparity boundaries. Besides, the proposed algorithm has low complexity and only requires a linear time implementation. Experimental results on Middlebury datasets demonstrate that the proposed method outperforms than other state-of-the-art methods.