Learning to render novel views from wide-baseline stereo pairs (CVPR2023)