Spatial deformable transformer for 3D point cloud registration.

Area of Science:

• Computer Vision
• Geometric Deep Learning
• 3D Data Processing

Background:

• Point cloud registration is crucial for 3D scene understanding.
• Traditional attention mechanisms can be computationally intensive and less effective for local geometric features.
• Extracting robust local geometric features is key to accurate point cloud registration.

Purpose of the Study:

• To propose a novel point cloud registration method, Spatial Deformable Transformer (SDT).
• To leverage deformable attention for efficient and accurate local feature extraction in point clouds.
• To improve matching recall, inlier ratio, and overall registration performance.

Main Methods:

• Developed Spatial Deformable Transformer (SDT) incorporating deformable self-attention and cross-attention modules.
• Deformable self-attention enhances local geometric feature representation.
• Cross-attention improves the discriminative capability of spatial correspondences.

Main Results:

• SDT demonstrates superior matching recall, inlier ratio, and registration recall on 3DMatch and 3DLoMatch datasets compared to state-of-the-art methods.
• The method exhibits better generalization ability and time efficiency on ModelNet40 and ModelLoNet40 datasets.
• SDT effectively captures dynamic local features without being constrained by feature map size.

Conclusions:

• Spatial Deformable Transformer (SDT) offers a significant advancement in point cloud registration.
• The proposed method achieves higher accuracy and efficiency by effectively utilizing deformable attention.
• SDT provides a robust and generalizable solution for various 3D registration tasks.