Abstract. Urban planning increasingly relies on structured 3D city models for analysis and decision-making, but while point clouds capture rich detail, their unstructured nature complicates direct use in CityJSON-based workflows. We address the challenge of translating raw point clouds into lightweight, interoperable models by confronting the lack of point connectivity, the difficulty of extracting stable planes, edges, and boundaries at scale, and the need for CityJSON-compliant outputs. We introduce a four-phase pipeline based on basic primitives (cuboid, pyramid, sphere, torus) that comprises specification and setup, data capture with EyesCloud3D, preprocessing through downsampling, noise removal, and normal estimation, and processing with RANSAC plane detection, DBSCAN clustering, Convex Hull and Alpha Shapes for boundary extraction, Triangle Normal Analysis for sharp-edge detection, volume-based RMSE evaluation, CityJSON conversion, and visualization in CityJSON Ninja. The approach yielded volume RMSEs of 309.14 cm³ for the cuboid, 70.2541 cm³ for the pyramid, 352.0292 cm³ for the sphere, and 58.1517 cm³ for the torus, while downsampling and denoising reduced point counts substantially, for example from 1,554,781 to 5,151 and then to 5,124 for the cuboid, and wireframes rendered efficiently and validated in CityJSON for the cuboid, pyramid, and sphere although the torus inner hole was not represented correctly. These results indicate that a plane and edge driven pipeline works best for sharp-edged, convex objects, with thresholds around 35 degrees for the cuboid and 30 degrees for the pyramid balancing sensitivity and noise, whereas performance degrades on smooth or non-convex shapes because Triangle Normal Analysis over detects tessellation artifacts on spheres and Convex Hull or Alpha Shapes cannot recover the torus topology. We conclude that the method is effective for sharp-edged urban forms and CityJSON-ready, and we suggest extending it with curvature-aware or implicit-surface reconstruction for smooth or non-convex geometries, adding semantic attributes, auto-tuning thresholds, and evaluating on compound building assemblies representative of real urban scenes.
Yee et al. (Tue,) studied this question.