Monocular Depth Estimation via Self-Supervised Self-Distillation.

Area of Science:

• Computer Vision
• Machine Learning
• Robotics

Background:

• Self-supervised monocular depth estimation excels in static scenes but struggles with dynamic environments due to occlusions.
• Maintaining depth consistency in dynamic scenes with moving objects presents a significant challenge.

Purpose of the Study:

• To propose a novel self-supervised self-distillation method for monocular depth estimation (SS-MDE) specifically designed for dynamic scenes.
• To enhance depth estimation accuracy in both static and dynamic regions within complex environments.

Main Methods:

• A deep network with a multi-scale decoder and lightweight pose network predicts depth using disparity and motion information.
• Pseudo-supervision from LeReS network refines static area depth, with a forgetting factor mitigating dependency.
• A teacher-student model framework with a multi-view mask filter enhances feature extraction and noise filtering.

Main Results:

• The SS-MDE method achieved superior performance on four public datasets compared to state-of-the-art techniques.
• Achieved 89% accuracy (δ1) and 0.102 AbsRel error on NYU-Depth V2.
• Attained 87% accuracy (δ1) and 0.111 AbsRel error on KITTI dataset.

Conclusions:

• The proposed SS-MDE method demonstrates enhanced generalization and robustness in dynamic scenes.
• Self-distillation effectively enables the student model to learn scene structure from a teacher model.
• The approach significantly advances self-supervised monocular depth estimation for real-world dynamic applications.