Abstract
Fine-grained 3D point cloud segmentation is essential for robot manipulation and CAD editing. A key unresolved challenge is that different parts of the same object often share similar local geometry, making part boundaries difficult to distinguish. Existing methods construct local patches via K-nearest search and rely on deeper and larger networks to learn discriminative features. However, these approaches mainly focus on local geometry and lack global structural awareness, which leads to ambiguous predictions under sampling noise and partial observations. In this work, we propose Topological Graph Attention Network (TopoGAT), a lightweight, plug-and-play backbone refinement module that integrates topological data analysis with Graph Attention Networks to introduce global structural information into point-wise feature learning. When combined with existing backbones, TopoGAT improves segmentation accuracy with a slight parameter increase. Extensive experiments on fine-grained 3D part segmentation validate the effectiveness of the proposed TopoGAT and show up to 1.0% improvement on ShapeNetPart dataset and 3.3% improvement on S3DIS dataset.