TIF-Reg: Point Cloud Registration with Transform-Invariant Features in SE(3).

Area of Science:

• Computer Vision
• 3D Reconstruction
• Robotics

Background:

• Point cloud registration (PCReg) is crucial for 3D applications but struggles with large-scale rigid transformations.
• Existing PCReg algorithms often fail to maintain accuracy under significant rotations and translations.

Purpose of the Study:

• To develop a robust point cloud registration method capable of handling large-scale rigid transformations.
• To improve the accuracy and efficiency of 3D point cloud registration.

Main Methods:

• Proposed TIF-Reg algorithm featuring transform-invariant feature extraction (TIF) in SE(3), deep feature embedding, attention-based point correspondence, and decoupled SVD.
• TIF incorporates triangular and local density features for enhanced transformation invariance.
• Deep neural networks and attention mechanisms are utilized for feature embedding and correspondence generation.

Main Results:

• TIF-Reg achieved root mean squared error (RMSE) for rotation within 0.5° and translation error near 0 m, even with transformations up to ±180° and ±20 m.
• The method demonstrated strong generalization capabilities on the TUM3D dataset after training on ModelNet40.
• Experimental results show superior accuracy and complexity compared to state-of-the-art PCReg algorithms.

Conclusions:

• TIF-Reg effectively addresses the challenge of large-scale rigid transformations in point cloud registration.
• The proposed method offers a robust and accurate solution for 3D computer vision and reconstruction tasks.
• TIF-Reg exhibits excellent performance and generalization, surpassing current state-of-the-art techniques.