Stereo Matching Using Multi-directional Dynamic Programming

In this paper, we propose a new stereo matching technique which employs an adaptive multi-path dynamic programming (DP) scheme. DP has been a classical and popular optimization method for various computer vision problems including stereo matching. However, the performance of conventional DP has not been satisfactory when applied to the stereo matching problems since the vertical correlation between scanlines has not been properly considered. This paper estimates disparity maps by using an optimal multi-directional DP scheme. We define a new energy function which considers the discontinuity of disparity and occlusions based on the edge information. The experimental results applied to the Middlebury stereo images demonstrate that our proposed algorithm shows better performances in stereo matching than the previous DP based approaches