3D Point Cloud Segmentation

The field studies how to divide unordered, irregular 3D points into meaningful classes, individual objects, or object parts, enabling machines to interpret spaces and physical structures. Methods include point- and voxel-based networks, sparse convolutions, graph models, and transformers, alongside clustering and prototype learning for unsupervised or few-shot settings. Researchers also combine point geometry with images, language models, Gaussian representations, and sensor attributes, and evaluate methods on benchmarks such as ScanNet, S3DIS, SemanticKITTI, and nuScenes. Key challenges include efficient processing of large scenes, precise boundaries, long-tailed classes, and reliable transfer across sensors, environments, and unfamiliar categories.