Self-supervised Monocular Depth Estimation and Ego-Motion Made Better: A Masking Constraints

In the field of depth estimation and visual odometry, self-supervised models based on monocular have gradually replaced LiDAR and stereo approaches. Self-supervised training typically relies on making assumptions about the brightness constancy and Lambertian object surfaces. However, if the assumptions within the frame are violated, the two modules in the self-supervised method will respectively experience noise and prediction errors. In this paper, we propose a masking constraints method to maintain the assumptions during training. Specifically, the region that breaks the assumptions is detected and extracted. Then, the region is converted into a mask for the self-supervised network. We combined our proposed method with representative models and tested the performance on deep prediction and visual odometry datasets provided by KITTI. The experiments show that the method can be integrated into self-supervised learning constraints, thereby reducing errors in depth estimation and ego-motion. The experiments also show that the method offers better interpretability and insights that can be derived from the learning process. Moreover, our analysis concludes that a reasonable reduction of pixel points improves the accuracy while the network learns the corresponding patterns faster.