Stereo matching based on segmented B-spline surface fitting and accelerated region belief propagation

The authors propose a new stereo matching algorithm based on an iterative optimisation framework including bi-cubic B-spline surface fitting and accelerated region belief propagation (BP). They first compute the initial cost and disparity map by the adaptive support-weight approach and then launch the iterative process in which the disparity space image is refined via the bi-cubic B-spline fitting and optimised via the accelerated region BP. Two innovations are contained in the algorithm: (i) disparity space image refinement based on segmented bi-cubic B-spline surface fitting; and (ii) an accelerated region message passing approach for BP. The algorithm is verified on the Middlebury benchmark and experimental results show the algorithm is effective and achieves the state-of-the-art accuracy.