Recovering viewer-centered depth from disparity, occlusion, and velocity gradients