Learning to refine depth for robust stereo estimation