关键词: 深度学习, 点云分类, 局部-全局交互, 结构Transformer

Abstract: Aiming at the problem of insufficient local and global feature extraction in the feature extraction process of point cloud classification, a point cloud classification algorithm with local-global interaction and structural Transformer is proposed. Firstly, a dual-branch parallel local-global interaction framework is proposed and extracts local and global features respectively, where one branch combines maximum pooling and convolution to extract local features, and the other branch extracts global features with average pooling and Transformer. Meanwhile, considering the importance of position information in Transformer, structural Transformer is proposed to further enhance global structural features by applying the interaction of position information with current features several times. Finally, the local-global features are used for classification to complete the classification task of the point cloud. Classification accuracies of 93.6% and 87.5% are obtained in ModelNet40 and ScanObjectNN benchmark datasets, respectively. The experimental results show that the proposed local-global interaction with structural Transformer network achieves good performance in the point cloud classification task.

Key words: deep learning, point cloud classification, local-global interaction, structural Transformer