Layered stereo matching using region boundary constraints

摘要

This paper presents a stereo matching algorithm that handles partial occlusion by exploiting region-boundary constraints. The novelty is that we use region boundaries as matching primitives, and define the global cost function by utilizing the photo-consistency and the smoothness constraints on the neighboring region boundaries. This algorithm is not only a reasonable way to locate the accurate occlusion parts of a segment, but it also can significantly reduce the number of vertex of a constructed graph to speed up the efficiency. Firstly, we use color information to segment the input image and obtain the initial disparity estimation by using a local matching algorithm. Then, initial disparity segments are clustered to form a set of disparity layers and assignments of region boundaries to disparity layers are iteratively optimized via a graph cuts framework. Finally, occlusion is detected based on the disparity layer assignments of the left and right region boundaries of the segments. The experimental results for the Middlebury test set justify the performance of our method.