3D Semantic Segmentation

3D semantic segmentation labels the parts of a three-dimensional scene, helping robots, vehicles, and mapping systems understand both what objects are and where they are. Common approaches process point clouds with point or voxel transformers, fuse LiDAR with camera features, or transfer 2D foundation-model predictions into 3D; newer work also segments Gaussian-splatting and mesh representations. Indoor benchmarks such as ScanNet and S3DIS and outdoor benchmarks such as SemanticKITTI and nuScenes measure progress, alongside few-shot, unsupervised, and open-vocabulary settings that reduce dependence on dense labels. A central challenge is making accurate boundaries and fine-grained predictions transfer across sensors, viewpoints, and domains without excessive annotation or computation.