The application of three-dimensional (3D) building models has expanded rapidly in recent years. Airborne lidar point clouds and optical images are two primary data sources for building reconstruction and provide complementary geometric and radiometric information. However, reliable 3D building reconstruction from the integration of both modalities remains challenging due to noise and occlusions in the data. To address these issues, this paper proposes a hypothesis-and-selection–based building reconstruction method that integrates airborne lidar point clouds with optical images using manually annotated image corner points. Roof structures are robustly extracted by an adaptive region-growing plane segmentation framework with iterative voting, which reduces undersegmentation near plane boundaries. The annotated image corners are then back-projected into object space. Facade positions are inferred according to their geometric relationships with the edge points of the point cloud. Finally, a compact and watertight polyhedral model is obtained by solving a binary linear programming problem. Bayesian optimization is introduced to automatically adjust the weights of energy terms and improve reconstruction stability. Experiments conducted on the public International Society for Photogrammetry and Remote Sensing Vaihingen and Toronto data sets demonstrate that the proposed method robustly produces geometrically accurate and structurally complete building models. Compared with existing methods, it achieves a better balance between geometric accuracy and model simplicity.
Wang et al. (Thu,) studied this question.