SC ²-PCR++: Rethinking the Generation and Selection for Efficient and Robust Point Cloud Registration.

Area of Science:

• Computer Vision
• Robotics
• 3D Data Processing

Background:

• Outlier removal is crucial for accurate feature-based point cloud registration.
• The classic RANSAC algorithm faces challenges with efficiency and robustness in complex scenarios.

Purpose of the Study:

• To improve model generation and selection in RANSAC for point cloud registration.
• To enhance the efficiency and robustness of outlier removal techniques.

Main Methods:

• Proposed a second-order spatial compatibility (SC²) measure for robust correspondence similarity assessment.
• Introduced a Feature and Spatial consistency constrained Truncated Chamfer Distance (FS-TCD) metric for model selection.
• Evaluated the generalizability of SC² and FS-TCD in deep learning frameworks.

Main Results:

• The SC² measure enables more effective inlier-outlier clustering with fewer samplings.
• The FS-TCD metric accurately selects the correct model even with very low inlier rates.
• Both proposed methods demonstrate improved performance in extensive experiments.

Conclusions:

• The SC² measure and FS-TCD metric significantly enhance outlier removal in point cloud registration.
• These methods offer improved efficiency and robustness compared to traditional approaches.
• The proposed techniques are versatile and compatible with deep learning-based registration frameworks.