MLDA-Net: Multi-Level Dual Attention-Based Network for Self-Supervised Monocular Depth Estimation.

Area of Science:

• Computer Vision
• Machine Learning
• Deep Learning

Background:

• Supervised learning for depth estimation requires extensive, costly annotations.
• Self-supervised methods are desirable but often produce blurry depth maps with lost details.

Purpose of the Study:

• To develop a novel framework, MLDA-Net, for self-supervised monocular depth estimation.
• To generate per-pixel depth maps with sharper boundaries and richer details.

Main Methods:

• Implemented a multi-level feature extraction (MLFE) strategy for hierarchical representation learning.
• Introduced a dual-attention strategy (global and structure attention) to enhance features.
• Utilized a reweighted loss strategy based on multi-level outputs for effective supervision.

Main Results:

• MLDA-Net achieves state-of-the-art results on the KITTI benchmark for self-supervised monocular depth estimation.
• Demonstrated superior performance across different input and training modes.
• Validated effectiveness on additional benchmark datasets.

Conclusions:

• MLDA-Net effectively addresses limitations of existing self-supervised depth estimation methods.
• The proposed framework produces significantly improved depth maps with enhanced detail and sharpness.
• MLDA-Net represents a significant advancement in self-supervised monocular depth estimation.