Multi-View Projection for Unsupervised Domain Adaptation in 3D Semantic Segmentation

3D semantic segmentation plays a pivotal role in autonomous driving and road infrastructure analysis, yet state-of-the-art 3D models are prone to severe domain shift when deployed across different datasets. In this paper, we propose an Unsupervised Domain Adaptation approach where a 3D segmentation model is trained on the target dataset using pseudo-labels generated by a novel multi-view projection framework. Our approach first aligns Lidar scans into coherent 3D scenes and renders them from multiple virtual camera poses to create large-scale synthetic 2D semantic segmentation datasets in various modalities. The generated datasets are used to train an ensemble of 2D segmentation models in point cloud view domain on each modality. During inference, the models process a large amount of views per scene; the resulting logits are back-projected to 3D with a depth-aware voting scheme to generate final point-wise labels. These labels are then used to fine-tune a 3D segmentation model in the target domain. We evaluate our approach Real-to-Real on the nuScenes and SemanticKITTI datasets. We also evaluate it Simulation-to-Real with the SynLidar dataset. Our contributions are a novel method that achieves state-of-the-art results in Real-to-Real Unsupervised Domain Adaptation, and we also demonstrate an application of our method to segment rare classes, for which target 3D annotations are not available, by only using 2D annotations for those classes and leveraging 3D annotations for other classes in a source domain.