Low Overlapping Point Cloud Registration Using Mutual Prior Based Completion Network.

Area of Science:

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

Background:

• Partial point cloud registration is crucial for 3D reconstruction and scene understanding.
• Existing methods struggle with low-overlapping point clouds due to insufficient common features.
• Lack of shape prior knowledge limits applicability across different point cloud types.

Purpose of the Study:

• To develop a novel completion method for low-overlapping partial point cloud registration.
• To enable registration without relying on external shape prior information.
• To improve the accuracy and robustness of point cloud registration in challenging scenarios.

Main Methods:

• Proposed Mutual Prior based Completion (MPC) method leverages candidate point clouds as mutual references.
• A progressive completion structure based on mutual information maximization is developed.
• Achieves joint optimization of pose, feature representation, and completion alignment.

Main Results:

• MPC successfully extends overlapping regions in low-overlapping partial point clouds.
• Demonstrated effectiveness across various point cloud types (objects, scenes, street views).
• Significant improvements over state-of-the-art methods: ~15.0% increase in overlap, 30.8% reduction in rotation error, and 57.7% reduction in translation error.

Conclusions:

• MPC offers a robust and effective solution for low-overlapping partial point cloud registration.
• The method's ability to work without shape priors broadens its applicability.
• Validated performance on public datasets shows substantial gains in registration accuracy.