Skip to Main Content
This paper provides a novel and efficient approach to dense stereo matching. The first contribution is, rather than applying disparity distribution inside a segment as hard constraint to directly project the likelihood of corresponding candidates to each pixel individually, our method treat segmentation and corresponding disparity distribution as soft constraint, and further partition each segment to sub-over segments which effectively facilitate the assumption of the disparity consistency. The second contribution is we transform this assumption into higher-order-based potential, and it can be minimized effectively through graph cut. The third contribution is the successful combination of several known techniques as one holistic framework. Two test-beds of both Middlebury and challenging real-scene data have been evaluated, results show that it obtains the state-of-the-art results while keeping efficiency.