GIPSO: Geometrically Informed Propagation for Online Adaptation in 3D LiDAR Segmentation

3D point cloud semantic segmentation is fundamental for autonomous driving. Most approaches in the literature neglect an important aspect, i.e., how to deal with domain shift when handling dynamic scenes. This can significantly hinder navigation capabilities of self-driving vehicles. paper advances state art this research field. Our first contribution consists analysing a new unexplored scenario segmentation, namely Source-Free Online Unsupervised Domain Adaptation (SF-OUDA). We experimentally show that state-of-the-art methods have rather limited ability adapt pre-trained deep network models unseen domains online manner. second approach relies on adaptive self-training and geometric-feature propagation source model without requiring either data or target labels. third study SF-OUDA challenging setup where synthetic clouds captured real world. use recent SynLiDAR dataset as introduce two (source) datasets, which stimulate future synthetic-to-real driving research. experiments effectiveness our thousands real-world (Code datasets are available at https://github.com/saltoricristiano/gipso-sfouda ).