Graph Convolutional Network using Self-Attention Module in Point Cloud Classification

Three-dimensional (3D) point cloud classification is one of the basic processing functions in deep learning to classify point cloud into a set of user-defined classes. This paper implements the latest state-of-the-art graph-based point cloud shape classification method-PointView-GCN and proposed a novel approach by integrating self-attention into graph convolution to enhance the classification accuracy of PointView-GCN. Self-attention is inspired by the Transformer architecture which assigns an associate weight for each updated feature after every local graph convolution. GCNNs benefits from this integration when the graph updates the features’ edges by selecting the most descriptive features among its k-nearest neighbourhood. The experimental results demonstrate that all variants of the integrated self-attention models consistently outperform the default PointView-GCN. Among the tested models, the best-performing architecture achieves an impressive accuracy of 93.08%, surpassing the default PointView-GCN by an improvement of approximately 1.49%. This notable enhancement in classification accuracy highlights the importance of integrating self-attention into PointView-GCN for point cloud analysis tasks.