Robust Point Cloud Registration via Rotation-Equivariant Geometric Encoding and State Space Models.

Area of Science:

• Computer Vision
• Robotics
• Geometric Deep Learning

Background:

• Point cloud registration is crucial for 3D scene understanding but struggles with textureless or repetitive environments.
• Existing methods face challenges in balancing local feature distinctiveness with global context modeling costs.
• Misalignments due to feature ambiguity and mismatch are common in challenging scenarios.

Purpose of the Study:

• To develop a robust and efficient point cloud registration framework for challenging environments.
• To mitigate feature ambiguity and mismatch issues inherent in current registration techniques.
• To achieve high accuracy and recall in point cloud registration tasks.

Main Methods:

• Proposed a framework combining rotation-equivariant geometric representations with linear-complexity state space models.
• Implemented a multivariate geometric encoding mechanism within convolutional layers for enhanced local features.
• Utilized a hybrid geometry-state aggregation module integrating self-attention with the Mamba architecture for long-range dependencies.
• Employed a physically consistent hypothesis generator for optimizing rigid transformation results.

Main Results:

• Achieved 96.3% registration recall on the 3DMatch dataset.
• Demonstrated outstanding accuracy on the KITTI dataset.
• Framework shows exceptional robustness to ambiguous matches in challenging environments.

Conclusions:

• The proposed method effectively addresses the limitations of traditional point cloud registration in difficult environments.
• Efficiently combines local feature enhancement with global context modeling using novel architectural components.
• Offers a robust and computationally efficient solution for accurate 3D point cloud registration.