Polyhedral representations with high-frequency for three-dimensional point cloud classification

Point clouds have become increasingly important due to their wide applications in fields such as autonomous driving and robotic navigation. However, the disordered and irregular nature of point clouds poses a challenge to traditional coordinate-based processing methods. These methods often struggle to accurately capture the geometric features of point clouds obtained in real environments, and the high-frequency structural information inherent in point clouds is often overlooked. In this study, a novel feature extraction method called Polyhedral Representations with High-Frequency (HF-Poly) is developed for point cloud classification. The HF-Poly method constructs tangent plane features for each point and creates polyhedral representations by neighboring tangent plane features. This approach enhances the capacity of the model to capture local features. Secondly, we design a novel high-frequency coding function that enhances the edge structure features using a high-frequency mapping strategy. This approach effectively delays the decay of these features in the deep network. In addition, we propose a general Local Channel Attention module (LCA), which assigns varying attention weights to different channels of the fused high-frequency polyhedral features to extract global information from the point cloud more efficiently. Extensive experiments on the ModelNet40 and ScanObjectNN datasets demonstrate that HF-Poly maintains superior classification accuracy while reducing computational costs compared to traditional methods, particularly achieving an impressive 88. 7 % accuracy on the ScanObjectNN dataset, with a 10. 8 % increase compared to PointNet++.

Authors

• Xiaoxin Mao
• Xue Li
• Puming Wang
• Xin Jin VCIP, CS, Nankai University
• Shengfa Miao
• Peng Liu
• Shaowen Yao School of Software and AI, Yunnan University
• Siwang Yang

Keywords

• Point cloud classification
• Polyhedral representation
• High-frequency encoding
• Local channel attention
• Deep learning