PointHGSA: Efficient Point Cloud Understanding with Hypergraph-Based Self-Attention Network

Zhou Jiang,Jing Yang,Jie Li,Dong Zhang,Shaoyi Du
DOI: https://doi.org/10.1109/itsc57777.2023.10422208
2023-01-01
Abstract:With the extensive research on autonomous driving and intelligent transportation technologies, the analysis and understanding of 3D point cloud data have become increasingly important. In this paper, we propose a point cloud understanding network called PointHGSA, targeting semantic segmentation and object classification tasks on point cloud data. The network introduces an innovative representation using a hypergraph structure, enabling a more flexible capture of complex relationships within the point cloud. This hypergraph structure considers relationships between multiple attributes simultaneously, resulting in a more accurate representation of semantic information in the point cloud. PointHGSA combines hyperedge convolution and self-attention mechanisms for feature encoding. Hyperedge convolution effectively considers the global relationships between points during feature propagation, providing a more comprehensive feature representation. The self-attention mechanism adaptively learns contextual information and captures the correlation between hyperedge and vertex features during feature adjustment, offering a more comprehensive modeling capability. Extensive experiments are conducted on the S3DIS, Semantic3D, and ModelNet40 datasets. The experimental results demonstrate that PointHGSA outperforms mainstream models on multiple datasets, particularly excelling in complex structures and objects with rich semantic information.
What problem does this paper attempt to address?