Integrating Spatial Context With Generative Model for Building Reconstruction From Airborne LiDAR Point Clouds

Reconstructing 3D buildings from airborne LiDAR point clouds is a fundamental task with critical importance in applications such as urban planning, autonomous driving, and emergency services. The emerging methods based on the generative model have achieved remarkable progress in this task. However, previous methods fail to fully exploit the spatial context inherent in point clouds, leading to inaccuracies in building reconstruction. This study proposes a novel method named SCGNet that integrates spatial context with the generative model to improve the reconstruction of 3D building models. Specifically, a global context-based vertex generation module is first introduced to predict building vertices. Then, a vertex-centric multi-scale spatial context encoding module is developed to capture the multi-scale semantics of spatial context that is centered on building vertices. Finally, a local context-enhanced face generation module is designed to reconstruct building faces. Extensive experiments demonstrate that the proposed method outperforms the baselines and is robust. This study explores a novel approach for optimizing building reconstruction, and provides a methodological reference for point-cloud-based 3D model tasks. The source code, model weights, and dataset will be available at: https://github.com/ShuaibZyx/SCGNet.