SC-DepthV3: Robust Self-Supervised Monocular Depth Estimation for Dynamic Scenes.

Area of Science:

• Computer Vision
• Machine Learning
• Deep Learning

Background:

• Self-supervised monocular depth estimation excels in static scenes but struggles with dynamic objects and occlusions.
• Existing methods fail in dynamic scenes, producing blurred depth maps at object boundaries due to occlusion in training views.

Purpose of the Study:

• To address the limitations of current self-supervised monocular depth estimation methods in dynamic scenes.
• To propose a novel approach, SC-DepthV3, for accurate depth map prediction in challenging dynamic environments.

Main Methods:

• Introduced an external pre-trained monocular depth estimation model to generate a single-image depth prior (pseudo-depth).
• Developed novel loss functions to leverage the pseudo-depth prior for boosting self-supervised training.
• Trained networks using monocular videos of highly dynamic scenes.

Main Results:

• SC-DepthV3 predicts sharp and accurate depth maps, even in highly dynamic scenes.
• The method significantly outperforms previous approaches on six challenging datasets.
• Detailed ablation studies validate the effectiveness of the proposed components.

Conclusions:

• SC-DepthV3 effectively overcomes the challenges posed by dynamic objects and occlusions in monocular depth estimation.
• The proposed pseudo-depth prior and novel losses enable robust and accurate depth prediction in complex real-world scenarios.