Building semantic segmentation from large-scale point clouds via primitive recognition

Modelling objects at a large resolution or scale brings challenges in the storage and processing of data and requires efficient structures. In the context of modelling urban environments, we face both issues: 3D data from acquisition extends at geographic scale, and digitization of buildings of historical value can be particularly dense. Therefore, it is crucial to exploit the point cloud derived from acquisition as much as possible, before (or alongside) deriving other representations (e.g., surface or volume meshes) for further needs (e.g., visualization, simulation). In this paper, we present our work in processing 3D data of urban areas towards the generation of a semantic model for a city digital twin. Specifically, we focus on the recognition of shape primitives (e.g., planes, cylinders, spheres) in point clouds representing urban scenes, with the main application being the semantic segmentation into walls, roofs, streets, domes, vaults, arches, and so on. Here, we extend the conference contribution in Romanengo et al. (2023a), where we presented our preliminary results on single buildings. In this extended version, we generalize the approach to manage whole cities by preliminarily splitting the point cloud building-wise and streamlining the pipeline. We added a thorough experimentation with a benchmark dataset from the city of Tallinn (47,000 buildings), a portion of Vaihingen (170 building) and our case studies in Catania and Matera, Italy (4 high-resolution buildings). Results show that our approach successfully deals with point clouds of considerable size, either surveyed at high resolution or covering wide areas. In both cases, it proves robust to input noise and outliers but sensitive to uneven sampling density.