Generating Watertight 3D Building Models from Airborne LiDAR Point Clouds using Detection Transformer (DETR)

Abstract. This work proposes a method for creating accurate, watertight 3D building models from airborne laser scanning (ALS) point clouds by leveraging a modified Detection Transformer (DETR) architecture. We adapted the DETR architecture to directly predict building planes from point clouds, from which the 3D model can be inferred using Boolean operations of half-spaces. We tested the model on the RoofN3D dataset and achieved a mean angle error of 1.7° for the building planes and a mean point-to-plane distance of 0.16m. Building facets can be detected even in the total absence of representative points, a common challenge in ALS data due to scanning direction and occlusions. By learning higher-level geometric principles, such as favouring 90-degree angles and symmetry, the model is able to adapt to various architectural styles without the need for explicit rules or pre-defined roof archetypes.