Misleading Supervision Removal Mechanism for self-supervised monocular depth estimation

Abstract

Self-supervised monocular depth estimation leverages the photometric consistency assumption and exploits geometric relations between image frames to convert depth errors into reprojection photometric errors. This allows the model train effectively without explicit depth labels. However, due to factors such as the incomplete validity of the photometric consistency assumption, inaccurate geometric relationships between image frames, and sensor noise, there are limitations to photometric error loss, which can easily introduce inaccurate supervision information and mislead the model into local optimal solutions. To address this issue, this paper introduces a Misleading Supervision Removal Mechanism(MSRM), aimed at enhancing the accuracy of supervisory information by eliminating misleading cues. MSRM employs a composite masking strategy that incorporates both pixel-level and image-level masks, where pixel-level masks include sky masks, edge masks, and edge consistency techniques. MSRM largely eliminate misleading supervision information introduced by sky regions, edge regions, and images with low viewpoint changes. Without altering network architecture, MSRM ensures no increase in inference time, making it a plug-and-play solution. Implemented across various self-supervised monocular depth estimation algorithms, experiments on KITTI, Cityscapes, and Make3D datasets demonstrate that MSRM significantly improves the prediction accuracy and generalization performance of the original algorithms.