A Review of Semantic Segmentation for Large-scale Point Cloud Data

To represent the real-world scene, a significant number of 3D points are utilized in 3D point cloud data. Comparatively, this data yields more information in higher dimensions compared to 2D image data. To effectively utilize the abundant information present in extensive scenes, machines must initially comprehend and recognize the meaning of individual 3D points. This understanding enables improved target identification and construction within these environments. Assigning specific meanings to each point in 3D point clouds is the process of semantic segmentation. This process classifies the entire point cloud into distinct categories, enabling important applications like scene robot navigation and autonomous driving. This paper presents a broad outline of the current datasets available for 3D point clouds and subsequently concentrates on summarizing the strategies employed to conduct semantic segmentation in extensive scenes. These techniques are classified into projection-based, discrete-based, and point-based methodologies, and inclusive discussions are provided for every category. Lastly, this paper highlights the current challenges in the field of research on 3D point cloud semantic segmentation.