Towards Robust Vision For Self-Supervised Depth And Motion Estimation