SGSLNet: Stratified Contextual Graph Pooling for Point Cloud Segmentation with Graph Structural Learning

Abstract

Recently, Graph Convolutional Neural Networks (GCNs) have demonstrated significant efficiency and flexibility in processing irregular data, exhibiting their considerable potential for point cloud segmentation. Because point cloud segmentation is essentially a point-wise classification task. However, current graph-based methods struggle to learn global structural outlines and local details effectively. Furthermore, the common application of Max pooling to aggregate point-wise features leads to a considerable loss of contextual information. To address these problems, we introduce a novel stratified graph structure learning network (SGSLNet). The main components of SGSLNet are adaptive structure-aware graph convolution (GAdaptive Conv) and stratified contextual graph pooling (SCGP). GAdaptive Conv is employed to learn local geometric structure dynamically, while SCGP applies to aggregate features and model global contextual structure. Our method not only learns global structural outlines and local details but also preserves substantial contextual information. We conduct extensive experiments on various datasets, including ShapeNetPart, S3DIS, and ScanNet v2. The results demonstrate that SGSLNet achieves state-of-the-art performance.