CGR-Block: Correlated Feature Extractor and Geometric Feature Fusion for Point Cloud Analysis

Area of Science:

• Computer Vision
• Machine Learning
• 3D Data Analysis

Background:

• Deep learning for point cloud processing is advancing rapidly.
• Existing networks struggle to extract both inter-feature interaction and geometric information simultaneously.
• Disordered 3D point clouds present unique challenges for feature extraction.

Purpose of the Study:

• To propose a novel point cloud analysis module, CGR-block, for enhanced feature extraction.
• To improve the simultaneous extraction of inter-feature interaction and geometric information.
• To develop a hierarchical network for point cloud classification and segmentation.

Main Methods:

• Introduction of the CGR-block module with two key units: correlated feature extractor and geometric feature fusion.
• Integration of a residual mapping branch within each CGR-block for performance enhancement.
• Construction of a classification and segmentation network using CGR-block as the fundamental module for hierarchical feature extraction.

Main Results:

• The proposed network achieved 94.1% overall accuracy on the ModelNet40 benchmark.
• The network attained 83.5% overall accuracy on the ScanObjectNN benchmark.
• An instance mIoU of 85.5% was achieved on the ShapeNet-Part benchmark, demonstrating superior performance.

Conclusions:

• The CGR-block module offers an efficient method for extracting geometric patterns and deep information from 3D point clouds.
• The developed network architecture significantly improves point cloud analysis tasks.
• The results validate the superiority of the proposed CGR-block based method over existing approaches.