SMNet: Surface Feature Extraction Network using Multi-scale Strategy

The current point cloud analysis methods primarily explore three-dimensional positional features, often overlooking other crucial surface features of 3D data, which are essential in 3D space. Additionally, most methods perform feature extraction at a single scale, neglecting the aggregation effects of variable neighborhoods at multiple scales. To address these limitations, we introduce SMNet, a network specifically designed for point cloud classification and segmentation tasks. SMNet generates additional surface features, such as curvature and normal vectors, using local positional information and updates these features globally through adaptive adjustments. In the overall network architecture, each feature extraction layer performs downsampling and uses different numbers of neighboring points to obtain neighborhoods, achieving a multi-scale feature extraction design. We evaluated SMNet on the ModelNet40 and ScanObjectNN datasets for point cloud shape classification, and on the ShapeNet Part dataset for point cloud part segmentation. The code will be released on https://github.com/NWUzhouwei/SMNet.