Improved Noise-driven Concurrent Stereo Matching Based on Symmetric Dynamic Programming Stereo

We describe a new version of the two-stage noisedriven concurrent stereo matching (NCSM) algorithm that uses symmetric dynamic programming stereo (SDPS) to estimate image noise and build candidate volumes for placing goal surfaces. In contrast to the initial NCSM, SDPS is first applied successively in xand y-directions in order to obtain a more stable disparity map from points that coincide in both cases and slant planes are fitted to the candidate volumes. This overcomes the principal shortcomings of dynamic programming techniques, such as propagation of matching errors along scan lines and reconstruction of a single continuous surface under the ordering constraint.