An efficient stereo matching method based on non-local spatial tree filter

Abstract

Binocular stereo vision is a vital research topic in computer vision and has been widely used in robot navigation, 3d reconstruction and other fields. Stereo matching is the most critical part of binocular stereo vision. In view of the low computational efficiency of local stereo matching methods, non-local methods based on tree structure have attracted much attention of researchers in recent years. In this paper, we propose a new efficient non-local spatial tree filter (NSTF) to aggregate the matching cost. Firstly, in addition to spatial affinity, the internal color similarity is taken as the similarity measure between adjacent pixels. Then, the propagation of cost is carried out recursively in the form of ternary tree. The whole filtering process can be divided into eight different directions. Quantitative experiments on Middlebury benchmark show that NSTF can effectively improve the accuracy of the algorithm, and has better edge-preserving ability than other tree-based non-local methods, especially in weak texture and high texture regions.