deep learning techniques have enabled rapid progress in monocular depth
estimation, but their quality is limited by the ill-posed nature of the problem
and the scarcity of high quality datasets. We estimate depth from a single
camera by leveraging the dual-pixel auto-focus hardware tha