Deep learning approaches for understanding 3D point clouds

Point clouds—sparse 3D data from sensors—present unique challenges due to their unordered structure, noise, and occlusions. This survey categorizes deep learning approaches for three core tasks: classification, part segmentation, and semantic segmentation. The authors organize methods by backbone architecture and evaluate them on standard benchmarks, covering strategies like format conversion, local geometry extraction, and permutation-invariant processing. The paper synthesizes performance comparisons, discusses architectural innovations and their limitations, and identifies open problems for future research.