Self-built dataset for better generalization in point cloud registration.

Area of Science:

• Computer Vision
• Machine Learning
• 3D Data Processing

Background:

• Point cloud registration networks often suffer from performance degradation and poor generalization due to training and testing datasets violating the independent and identically distributed (IID) assumption.
• Existing methods struggle to create datasets that adequately represent real-world domain discrepancies.

Purpose of the Study:

• To address the IID assumption violation in point cloud registration datasets.
• To develop a post-processing method for generating diverse point cloud data.
• To improve the accuracy and generalization of point cloud registration networks.

Main Methods:

• Proposed a novel point cloud data post-processing technique.
• Introduced domain discrepancy variables by optimizing spatial sampling and temporal interval frame matching.
• Generated a large dataset of 63,461 point cloud registration data pairs.

Main Results:

• Achieved a 31.8% improvement in accuracy compared to benchmark datasets.
• The one-step generalization ratio reached 0.9832, indicating significantly enhanced network generalization.
• The generated dataset provides valuable reference data for cross-domain point cloud registration research.

Conclusions:

• The proposed post-processing method effectively addresses the IID assumption in point cloud registration.
• The enhanced dataset leads to substantial improvements in accuracy and generalization capabilities of registration networks.
• This work offers a robust solution and valuable resources for advancing cross-domain point cloud registration research.