SDA-Net: A Global Feature Point Cloud Completion Network Based on Serialization and Dual Attention.

Area of Science:

• Computer Vision
• 3D Geometry Processing
• Machine Learning

Background:

• Point cloud completion is crucial for 3D data restoration, but current methods struggle with global structural information.
• K-nearest neighbor (KNN) methods focus locally, while Transformer methods often use windowed attention, limiting global context.
• Existing approaches fail to fully capture the intricate relationships within unordered point cloud data.

Purpose of the Study:

• To introduce SDA-Net, a novel dual-attention network for point cloud completion.
• To address the limitations of existing methods in modeling global context and inter-point relationships.
• To improve the accuracy and detail of reconstructed 3D point clouds.

Main Methods:

• Developed SDA-Net, a dual-attention network incorporating multiple serialization strategies.
• Transformed unordered point clouds into structured sequences to model inter-point relationships comprehensively.
• Employed a dual-attention mechanism with spatial and channel-wise self-attention for enhanced global feature extraction.

Main Results:

• SDA-Net achieved state-of-the-art performance with an average Chamfer Distance (CD) of 6.48 on the PCN dataset.
• Demonstrated superior accuracy in reconstructing fine-grained details for real-world LiDAR-scanned point clouds.
• Effectively compensated for the loss of global context in point cloud completion tasks.

Conclusions:

• SDA-Net offers a robust solution for point cloud completion by effectively integrating global and local features.
• The proposed dual-attention and serialization strategies significantly advance the state-of-the-art in 3D geometric data restoration.
• SDA-Net shows strong potential for practical applications requiring high-fidelity 3D reconstruction.