Segmentation-based adaptive support for accurate stereo correspondence