Parallel Multi-Scale Semantic-Depth Interactive Fusion Network for Depth Estimation.

Area of Science:

• Computer Vision
• Machine Learning

Background:

• Self-supervised depth estimation from monocular images is crucial for autonomous driving, reducing reliance on expensive sensors like LiDAR.
• Current methods struggle with occlusions, lighting variations, and sparse textures, and lack feature enhancement and fusion.
• Existing approaches do not adequately integrate semantic and depth information for improved accuracy.

Purpose of the Study:

• To propose a novel parallel multi-scale semantic-depth interactive fusion network for enhanced self-supervised depth estimation.
• To address limitations in feature enhancement and interaction fusion present in existing methods.
• To improve the accuracy and robustness of depth estimation in challenging driving scenarios.

Main Methods:

• Utilized a multi-stage feature attention network for robust feature extraction.
• Introduced a parallel semantic-depth interactive fusion module to refine object edges and details.
• Employed a metric loss based on semantic edges to leverage geometric information.

Main Results:

• The proposed network demonstrated satisfactory performance on the KITTI dataset.
• Achieved improved depth estimation accuracy compared to existing self-supervised methods.
• The fusion of semantic and depth features effectively enhanced edge refinement and overall performance.

Conclusions:

• The novel network effectively integrates multi-scale features and semantic information for superior self-supervised depth estimation.
• The proposed approach offers a promising solution for enhancing depth perception in autonomous driving systems.
• Future work can explore further integration of contextual information and advanced attention mechanisms.