Constructing Diverse Inlier Consistency for Partial Point Cloud Registration.

Area of Science:

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

Background:

• Partial point cloud registration aligns incomplete 3D scans.
• Existing methods struggle with intra- and inter-point cloud relationships.
• Accurate correspondence identification and overlap detection are challenging.

Purpose of the Study:

• To propose a novel Diverse Inlier Consistency (DIC) method for partial point cloud registration.
• To enhance the utilization of relative positional information within and between point clouds.
• To improve the accuracy of overlapping region identification and correspondence searching.

Main Methods:

• Developed a Diverse Inlier Consistency-driven Region Perception (DICdRP) module for intra-point cloud positional encoding.
• Introduced a Diverse Inlier Consistency-aware Correspondence Search (DICaCS) module for inter-point cloud relative positioning.
• Integrated diverse information for a comprehensive registration framework.

Main Results:

• The proposed DIC method demonstrates superior performance on object-level and scene-level datasets.
• DIC effectively utilizes intra- and inter-point cloud positional information.
• Enhanced accuracy in identifying overlapping regions and filtering outliers.

Conclusions:

• The DIC method offers a significant advancement in partial point cloud registration.
• The approach effectively addresses limitations of previous learning-based methods.
• The framework provides a more holistic and detailed registration process.