GAC-Net: A Geometric-Attention Fusion Network for Sparse Depth Completion from LiDAR and Image.

Area of Science:

• Computer Vision
• Robotics
• Machine Learning

Background:

• Depth completion reconstructs dense depth maps from sparse LiDAR and RGB data.
• Existing methods like BPNet show promise but can be improved with better 3D geometric priors and adaptive fusion.

Purpose of the Study:

• To propose GAC-Net, a novel Geometric-Attention Fusion Network.
• Enhance geometric representation and cross-modal fusion for depth completion.

Main Methods:

• A dual-branch PointNet++-S encoder extracts scale-aware geometric features from sparse point clouds.
• Channel attention fuses these features for a robust 3D representation.
• A Channel Attention-Based Feature Fusion Module (CAFFM) adaptively integrates geometric, RGB, and depth features.

Main Results:

• GAC-Net achieved a Root Mean Square Error (RMSE) of 680.82 mm on the KITTI depth completion benchmark.
• This performance ranked first among peer-reviewed methods at the time of submission.

Conclusions:

• GAC-Net effectively enhances geometric representation and cross-modal fusion for depth completion.
• The proposed network demonstrates superior performance on a standard benchmark, advancing the field.