Self-supervised Pretraining and Finetuning for Monocular Depth and Visual Odometry