A fast non-local disparity refinement method for stereo matching

This paper presents a fast non-local disparity refinement method based on disparity belief propagation. The disparity belief fast propagated on a minimum spanning tree only need two sequential passes, first from leaf nodes to root, then from root to leaf nodes. Computational complexity of each pixel at all disparity levels is O(1). Performance evaluation on standard Middlebury data sets shows that the proposed method outperforms local refinement method both in accuracy and speed. Compared with the existing nonlocal disparity refinement method, the proposed method shows about maximum 15× faster speed at almost the same accuracy.