The Multi-View Deep Visual Adaptive Graph Convolution Network and Its Application in Point Cloud

• Published in: Traitement du signal 2023-02, Vol.40 (1), p.31-41
• Main Authors: Fan, Haoyang, Zhao, Yanming, Su, Guoan, Zhao, Tianshuai, Jin, Songwen
• Format: Article
• Language: English
• Subjects: Algorithms; Classification; Convolution; Deep learning; Graph representations; Invariance; Machine learning; Methods; Recognition; Three dimensional models
• ISSN: 0765-0019; 1958-5608
• DOI: 10.18280/ts.400103

Regarding the classification of 3D point clouds, existing Graph Convolution Networks (GCN) often fail to effectively learn the correlation of visual features of different scales under the condition of multi-view (multi-domain), thus the features learnt by models are not as varied and the classification accuracy is usually limited. In view of these defects, this paper proposed a Multi-View Deep Visual Adaptive Graph Convolution Network (MVDVAGCN) which integrates the theory of visual selective attention with the graph convolution calculation method and inherits the advantages of the idea of rasterization. In the paper, the deep learning technology, the idea of multi-view, and the VAGCN were combined to establish three GCN models which were then applied to the classification of 3D point clouds and attained good results. Then the parameters set for the proposed algorithm were verified based on the ModelNet40 dataset, the recognition performance and geometrical invariance of the proposed algorithm and a few reference algorithms including VoxNeT, PointNet, PointNet++, DGCNN, KPConv3D-GCN, Dynamic Graph CNN, and 3D_RFGCN, were tested, and the results proved the feasibility, recognition performance, and geometrical invariance of the proposed algorithm.