Non-local affinity adaptive acceleration propagation network for generating dense depth maps from LiDAR.

Area of Science:

• Computer Vision
• Machine Learning
• Robotics

Background:

• Depth completion is crucial for generating dense depth maps from sparse LiDAR data.
• Existing methods struggle with depth boundaries, leading to mixed-depth issues for adjacent objects.

Purpose of the Study:

• To propose a novel Non-Local Affinity Adaptive Accelerated (NL-3A) propagation network for robust and efficient depth completion.
• To address the challenge of accurately reconstructing depth at object boundaries.

Main Methods:

• Designed an NL-3A prediction layer for initial depth maps, pixel reliability, non-local neighbors, affinities, and learnable normalization factors.
• Developed an NL-3A propagation layer integrating non-local neighbor affinity with pixel reliability for adaptive propagation weights.
• Implemented an accelerated propagation model enabling parallel processing for enhanced efficiency.

Main Results:

• The NL-3A network outperforms existing algorithms in accuracy and efficiency on KITTI and NYU Depth V2 datasets.
• Demonstrated superior performance in predicting and reconstructing depth smoothly and consistently at object edges.
• Overcame propagation errors associated with mixed-depth objects through non-local neighbor prediction.

Conclusions:

• The proposed NL-3A network offers a significant advancement in depth completion technology.
• Achieved state-of-the-art results in accuracy, efficiency, and boundary reconstruction for dense depth maps.
• The adaptive and accelerated propagation mechanism enhances network robustness and practical applicability.