Lightweight 3D Point Cloud Classification Network

In this paper, we propose a new lightweight neural network for point cloud classification that has only about 100 K parameters. Most of the current research focuses on aggregating network features through pooling layers and extracting abstract features of 3D point clouds using higher dimensions. In this work, we turn our attention to exploring the relationships between points at a deeper level, using a multilayer perceptron module with a residual structure to suppress the performance degradation problem that comes with increasing the number of network layers. On the ModelNet40 dataset, our method achieves an accuracy of more than 92.5%, which is the first of its kind in our knowledge for ultralight networks. Without using any techniques such as pruning and quantization, the model was trained at 1214 samples per second and inferred at a staggering 1956 samples per second.