A Comprehensive Survey on Deep Learning‐Based Semantic Segmentation of 3D Point Clouds: Technologies, Challenges, and Future Directions

ABSTRACT

3D point cloud semantic segmentation is a fundamental task in 3D scene understanding and has become increasingly important in applications such as robotics, autonomous driving, and immersive environments. Although existing surveys have reviewed representative architectures and datasets, they often focus on closed‐set geometric segmentation and provide limited framework‐level analysis of recent developments such as Transformer‐based modelling, multimodal fusion, 3D foundation‐model‐related learning, open‐vocabulary segmentation, and language‐assisted perception. This survey revisits deep learning‐based 3D point cloud semantic segmentation through a consistent multi‐axis taxonomy covering task scope, supervision regime, representation/backbone design, modality configuration, and semantic openness. We review representative methods, clarify the boundary between semantic segmentation and adjacent 3D understanding tasks, and summarise benchmark datasets, evaluation protocols, and comparability caveats to support more reliable interpretation of reported results. We further discuss major challenges, including annotation efficiency, multimodal misalignment, 2D–3D semantic gaps, open‐world generalisation, and computational scalability, and outline future directions towards more robust, transferable, and deployable 3D perception systems.