NPMFF-Net: A training-free unified framework for point cloud classification and segmentation

Abstract

Non-parametric networks have shown promise for understanding point clouds due to their training-free nature and low computational cost. However, existing methods such as Point-NN and Seg-NN underutilize geometric and frequency information. Although these methods demonstrate superior accuracy, we found that the potential features of point clouds can still be explored in depth. In this work, we revisit non-parametric networks and propose the Non-Parametric Multi-scale Feature Fusion Network (NPMFF-Net), a model designed to unify spatial and frequency information in point cloud analysis, featuring training-free components. The key is Plücker coordinates Encoding and Fourier Feature Mapping, combining geometric information with high-frequency features. We propose a non-parametric attention module to integrate contextual information and k-adaptive normal pooling to aggregate multi-scale features. Extensive experiments on the ModelNet10/40, ScanObjectNN, ShapeNetPart, S3DIS, and ScanNet datasets demonstrate the superiority of NPMFF-Net in point classification and segmentation tasks. We surpass Point-NN by 8.2 % OA and Seg-NN by 5.8 % OA on ModelNet40 for classification, while also achieving a 2.7 % improvement in mean IoU over Point-NN on ShapeNetPart for part segmentation.