PointGS: Unsupervised 3D Point Cloud Segmentation via Gaussian Splatting

Researchers propose PointGS, a novel pipeline for unsupervised 3D point cloud segmentation that uses 3D Gaussian Splatting as an intermediate representation to bridge the gap between discrete 3D points and continuous 2D images. The method addresses the mismatch that occurs when integrating 2D pre-trained models like SAM, which leads to projection overlap and complex modality alignment. PointGS first reconstructs input sparse point clouds into dense 3D Gaussian spaces, enabling semantic-consistent segmentation without requiring dense point-level annotations. The approach is critical for embodied AI and autonomous driving, where annotation costs are prohibitive. The paper is available on arXiv under ID 2605.11520.