LogoSP: Local-global Grouping of Superpoints for Unsupervised Semantic Segmentation of 3D Point Clouds

Input Point Cloud GrowSP PointDC PointDC-DINOv2 LogoSP (Ours) Ground Truth Figure 1. Given an input point cloud with complex structures from ScanNet dataset, our LogoSP automatically discovers accurate semantic classes such as sofa and table without needing human labels in training, clearly outperforming all existing baselines.