Digging Into Self-Supervised Monocular Depth Estimation