Abstract
Depth completion aims to reconstruct dense depth maps from sparse LiDAR measurements guided by RGB images. Although BPNet enhanced depth structure perception through a bilateral propagation module and achieved state-of-the-art performance at the time, there is still room for improvement in leveraging 3D geometric priors and adaptively fusing heterogeneous modalities. To this end, we proposed GAC-Net, a Geometric-Attention Fusion Network that enhances geometric representation and cross-modal fusion. Specifically, we designed a dual-branch PointNet++-S encoder, where two PointNet++ modules with different receptive fields are applied to extract scale-aware geometric features from the back-projected sparse point cloud. These features are then fused using a channel attention mechanism to form a robust global 3D representation. A Channel Attention-Based Feature Fusion Module (CAFFM) was further introduced to adaptively integrate this geometric prior with RGB and depth features. Experiments on the KITTI depth completion benchmark demonstrated the effectiveness of GAC-Net, achieving an RMSE of 680.82 mm, ranking first among all peer-reviewed methods at the time of submission.