Stereo Matching by RBF Networks with Occlusion Detection

In this paper, we propose a new solution to the stereo correspondence problem by using both features and intensity values. A cost function is formed by integrating intensity errors, gradient errors, and a smoothness error. Unlike previous use of a smoothness constraint, the smoothness term here is applied only to non-feature regions. To do function minimization, we propose to parameterize the disparity function by hierarchical Gaussians. To detect occlusions and disparity discontinuities, we propose to combine results from the left-to-right and right-to-left matchings without explicitly modeling discontinuities in the algorithm. Experimental results show efficiency of our method.