PC2Mesh: Reconstruction of building polygonal meshes from airborne LiDAR point clouds

Polygonal meshes provide a compact and flexible representation for 3D buildings, supporting a wide range of photogrammetric and geometric applications. However, reconstructing high-fidelity building polygon meshes from airborne LiDAR point clouds remains challenging due to the non-uniform and variable-sided nature of polygonal topology and the complexity of urban environments characterized by diverse building geometries, uneven point densities, and occlusions. In this paper, we present PC2Mesh, a learning-based method that reconstructs polygonal building meshes directly from airborne LiDAR point clouds. While existing approaches rely on primitive detection (e.g., City3D, Point2Poly) or employ separate networks for vertex and face generation (e.g., Point2building), our method employs a unified transformer-based architecture that jointly learns mesh geometry and topology within a shared representation space. By representing arbitrary-sided polygons as learned token sequences, the method enables holistic mesh generation without preprocessing or handcrafted priors. This end-to-end formulation mitigates error propagation across stages and enhances reconstruction fidelity and robustness. Comprehensive experiments on open-source datasets demonstrate that our proposed method consistently outperforms state-of-the-art approaches in geometric accuracy and visual fidelity. Our method improves Edge Precision (EP) and Edge Recall (ER) by 9.1% and 24.7%, respectively, on the Zurich dataset, and by 8.3% and 20.2% on the Tallinn dataset. It also achieves the lowest Mean Distance Error (MDE), Hausdorff Distance (HD), and Chamfer Distance (CD) on both datasets. The source code of this work is freely available at https://github.com/BestyHon/PC2Mesh .