GPNN:Grouped Point-NN for more efficient point cloud classification

Point cloud is irregular and permutation-invariant, unlike grid-based data such as images, posing unique challenges for effective analysis and processing. PointNN, proposed by Zhang et al., introduces a non-parametric network for point cloud analysis, achieving competitive performance on multiple 3D tasks without any learnable parameters or training. However, the pooling operation in the model can lead to a sharp reduction in the number of features, resulting in a considerable loss of important information. To solve this issue, we introduce an enhancement mechanism termed "Subset Division(SD)" into the original PointNN. "Subset Division" is a simple mechanism where the input point cloud is divided into multiple subsets before being fed into the model. Despite the simple mechanism of "Subset Division" can enhance classification performance, and "PointNN+SD" achieves a higher classification performance of 82.2% on ModelNet40 compared to the original PointNN. PointPN adds a few learnable linear layers to the base model PointNN and improves its classification accuracy by incorporating "subset division." Additionally, by ensembling the output results of "PointNN+SD" and the original, we achieve an even greater classification accuracy of 83.7%, higher than either model individually.