One-Nearest Neighborhood Guides Inlier Estimation for Unsupervised Point Cloud Registration.

Area of Science:

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

Background:

• Unsupervised point cloud registration accuracy is hindered by unreliable inlier estimation and self-supervised signals, particularly with partial overlaps.
• Existing methods struggle to robustly identify correct correspondences in challenging registration scenarios.

Purpose of the Study:

• To develop an effective inlier estimation method for unsupervised point cloud registration.
• To improve registration precision by capturing geometric structure consistency.
• To provide reliable self-supervised signals for unsupervised model optimization.

Main Methods:

• Generated a high-quality reference point cloud copy using a one-nearest neighborhood (1-NN) approach.
• Integrated dual neighborhood matching scores (1-NN and input point cloud) to enhance matching confidence.
• Constructed transformation-invariant geometric structure representations to score inlier confidence based on neighborhood graph consistency.
• Employed a weighted Singular Value Decomposition (SVD) algorithm for transformation estimation.

Main Results:

• The proposed method demonstrates effective inlier estimation by exploiting geometric structure consistency.
• Dual neighborhood matching significantly improves matching confidence and registration accuracy.
• Transformation-invariant representations provide reliable self-supervised signals for unsupervised training.
• Experiments on synthetic and real-world datasets validate the method's effectiveness.

Conclusions:

• The proposed unsupervised point cloud registration method achieves high precision through robust inlier estimation.
• The strategy of capturing geometric structure consistency offers a powerful self-supervised signal for registration.
• This approach effectively addresses limitations in partially overlapping point cloud registration.