SRNSD: Structure-Regularized Night-Time Self-Supervised Monocular Depth Estimation for Outdoor Scenes.

Area of Science:

• Computer Vision
• Machine Learning
• Robotics

Background:

• Deep Convolutional Neural Networks (CNNs) excel at daytime monocular depth estimation.
• Night-time conditions (low visibility, varying illumination) significantly degrade performance due to domain gaps.

Purpose of the Study:

• To develop a novel framework for robust night-time self-supervised monocular depth estimation.
• To address performance degradation caused by day-night domain gaps and challenging illumination.

Main Methods:

• Proposed a framework with structure regularization (SRNSD) incorporating feature/depth domain adaptation.
• Utilized high- and low-frequency decoupling for structure and texture recovery.
• Implemented image perspective constraint and cropped multi-scale consistency loss for enhanced depth prediction.

Main Results:

• Demonstrated superior performance over state-of-the-art methods on Oxford RobotCar, nuScenes, and CARLA-EPE datasets.
• Achieved significant improvements in night-time depth estimation accuracy across multiple metrics.
• Validated effectiveness on depth ranges up to 60m.

Conclusions:

• The proposed SRNSD framework effectively overcomes challenges in night-time monocular depth estimation.
• Structural regularization and multi-scale consistency are crucial for accurate night-time depth prediction.
• The approach offers a significant advancement for autonomous systems operating in low-light conditions.