RIGA: Rotation-Invariant and Globally-Aware Descriptors for Point Cloud Registration

Area of Science:

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

Background:

• Accurate point cloud registration is crucial for 3D data analysis.
• Existing neural descriptors struggle with large rotations and lack distinctiveness.
• Rotation-variant backbones and limited local geometry encoding hinder performance.

Purpose of the Study:

• To develop novel rotation-invariant and globally-aware descriptors for enhanced point cloud registration.
• To address the limitations of current neural descriptors in handling large rotations and capturing global context.
• To improve the accuracy and robustness of point cloud registration.

Main Methods:

• Introduced RIGA (Rotation-Invariant and Globally-Aware) descriptor learning framework.
• Encoded rotation-invariant local geometry using Point Pair Features (PPFs).
• Incorporated global awareness via global PPF signatures and aggregated geometric context.
• Interpolated sparse region descriptors to dense point descriptors for correspondence extraction.

Main Results:

• RIGA surpasses state-of-the-art methods in relative rotation error on ModelNet40 by 8 degrees.
• Achieved at least a 5 percentage point improvement in Feature Matching Recall on 3DLoMatch.
• Demonstrated superior performance on both object- and scene-level point cloud registration tasks, especially under large rotations.

Conclusions:

• RIGA provides robust and accurate point cloud registration, particularly in challenging scenarios with large rotations.
• The proposed method effectively combines local geometric information with global structural context.
• RIGA represents a significant advancement in learning powerful and invariant descriptors for 3D point cloud analysis.