AirDC: Adaptive Iterative Depth Refinement Framework for Full-Range Metric Depth Completion.

Area of Science:

• Computer Vision
• Robotics
• Sensor Fusion

Background:

• Autonomous systems require accurate metric depth completion across diverse ranges.
• Existing methods struggle with fine-grained depth feature capture at both near and far distances.
• Inadequate modeling of range-specific depth cues limits current depth completion techniques.

Purpose of the Study:

• To propose AirDC, an adaptive iterative depth refinement framework for full-range metric depth completion.
• To improve the efficient capture of depth features across wide depth ranges.
• To address limitations in modeling fine-grained depth cues for enhanced autonomous system perception.

Main Methods:

• Developed an adaptive fine-grained stereo-LiDAR feature fusion module utilizing metric-aligned depth volumes and sub-voxel depth attention.
• Introduced an iterative hypothesis-guided depth refinement module integrating multi-scale, multi-modal guidance.
• Employed an adaptive sub-voxel depth attention mechanism for enhanced sensitivity to subtle depth variations.

Main Results:

• AirDC demonstrates state-of-the-art performance on multiple real-world and synthetic benchmarks.
• The framework effectively preserves original full-range depth information.
• Achieved improved prediction accuracy with memory efficiency and a small parameter overhead.

Conclusions:

• AirDC provides an effective solution for full-range metric depth completion.
• The proposed adaptive fusion and iterative refinement modules significantly advance depth estimation capabilities.
• The framework offers a robust approach for enhancing perception in autonomous systems.