Applying the Hungarian method to stereo matching

The correspondence problem in stereo vision is to calculate matches between pixels (points) or features (e.g., lines) in stereo images. The method of calculating these matches over an entire stereo image pair usually depends upon the application for which the resulting depth information will be used. Functional minimization and dynamic programming are two widely used techniques for selecting matches. We present an alternative technique based on integer-programming formulation of the correspondence problem. Computation of the disparity map is accomplished in polynomial time by casting the selection of matches along corresponding epipolar lines as the problem of finding a maximum weighted matching on bipartite graphs. This approach incorporates several desirable properties of disparity maps, including implicit occlusion estimation, preference for fronto-parallel surfaces, and improved performance in the absence of texture (when compared to non-global techniques which do not maximize over the entire epipolar line).