A lightweight point cloud analysis network based on graph walking

Based on the irregularity and permutation invariance of point clouds, graph neural networks have been widely used in 3D point cloud analysis tasks in recent years. However, these works suffer from insufficient feature extraction at long distances and high computational resource overhead. To this end, this paper proposes a graph walk based lightweight network for 3D point cloud analysis tasks. Specifically, after constructing the nearest neighbor graph in the downsampled point cloud, the MLP module in the original network was replaced by multiple nonparametric components to reduce the number of parameters of the model. Then, the graph walking mechanism was used to extract the long-distance features and geometric structure features of the point cloud. Finally, before point classification and segmentation, the residual pooling module is used to fuse the features to obtain the global features. Compared with the benchmark network DGCNN, the classification accuracy on the ScanObjectNN dataset is improved by 4.2%, the segmentation accuracy on the ShapeNetPart dataset is improved by 1.1%, and the number of model parameters is reduced by 33.3%. Experiments show that the design method has obvious effect on 3D point cloud analysis tasks.