Large-scale 3D Mesh Data Semantic Segmentation: A Survey

Utilizing 3D mesh data presents myriad advantages for representing actual environments, encompassing the capability to exhibit geometric information and high-quality textures. Within the domain of geographical information systems, there exists a substantial need for 3D data to precisely depict extensive remote sensing scenes. Meshes are commonly employed for representing geographical terrains, structures, and urban models. Effective visualization and representation of meshes are imperative for managing expansive geographical datasets. The procedure of 3D mesh semantic segmentation entails assigning distinct semantic labels to each triangle, thus dividing the entire scene into various categories. Nevertheless, owing to the irregular structure of 3D mesh data, the progress in developing pertinent semantic segmentation techniques is still in its nascent stages. Many conventional methods rely on multi-view approaches. Given the growing popularity of deep learning for segmenting remote sensing data, end-to-end 3D mesh semantic segmentation techniques have become a focal point of current research. This article firstly introduces existing 3D mesh datasets and classifies them based on scale and application, as these datasets serve as the groundwork for algorithmic research. Subsequently, existing 3D mesh semantic segmentation techniques are presented, categorized into traditional methods and deep learning methods. Lastly, the challenges confronted by 3D mesh semantic segmentation at present are deliberated.