Główne pojęcia
Our method produces simple, regularized 3D building models that are faithful approximations of the input LiDAR data, while being efficient and offering strong geometric guarantees.
Streszczenie
The authors propose a method called SimpliCity for reconstructing 3D building models from airborne LiDAR point clouds. The key ideas are:
Construction of a 2D polygonal partition that represents the roof structure:
Detect 3D planes from the input point cloud and construct a 2D polygonal partition by projecting intersection and discontinuity lines.
Assign a label to each cell of the partition corresponding to one of the detected planes or the ground.
Regularization of the 2D polygonal partition:
Collapse short edges to simplify the partition.
Detect and enforce near-parallel and near-orthogonal constraints between edges to enhance regularity.
Optimize the vertex positions globally under these geometric constraints.
Extrusion of the regularized 2D partition to 3D:
Extrude each cell of the partition to 3D using the associated plane equations.
Optimize the vertex heights to ensure continuity between adjacent roof sections while preserving planarity.
The authors evaluate their method on three datasets and compare it to several baselines. Their approach produces significantly simpler 3D building models (2-4 times fewer vertices and facets) while maintaining similar accuracy and efficiency compared to prior methods. The reconstructed models also exhibit strong geometric guarantees, being watertight, 2-manifold, and intersection-free.
Statystyki
The average number of vertices |V| in the reconstructed models is 32.7, 58.3, and 214 for the Tallinn, Zurich, and Helsinki datasets, respectively.
The average number of facets |F| in the reconstructed models is 58.2, 109, and 393 for the Tallinn, Zurich, and Helsinki datasets, respectively.
The ratio of edges shorter than 0.5 meters E<0.5m in the reconstructed models is 4.39%, 7.48%, and 9.5% for the Tallinn, Zurich, and Helsinki datasets, respectively.
Cytaty
"Our algorithm offers the most desired guarantees with watertight, 2-manifold, intersection-free meshes."
"We show the benefits of our approach against prior methods by producing more simple 3D models while reaching a similar fidelity and efficiency."