Automatic Generation of LOD3 Building Models for High-Density Cities: A Case Study of Hong Kong Using Multi-Source Data and an Adaptive Strategy

Abstract. This paper presents a framework for upgrading LOD2 building models to LOD3 by directly exploiting texture information embedded in existing LOD2 models. A semi-automated annotation tool based on SAM2 enables instance-level segmentation of facade components from texture images, with 2D bounding boxes back-projected to 3D coordinates via UV mapping and barycentric interpolation. Recognized textures serve as queries for two parallel pipelines: metric learning with DINOv2 retrieves similar components from a curated BIM library, while TripoSR generates 3D components directly from single texture images. All retrieved or generated components are adaptively scaled and fused with the LOD2 model using Boolean operations based on constructive solid geometry. Experimental results demonstrate that the metric learning approach achieves 81.2% Precision@1 and 46.6% MAP@k with SubCenter ArcFace loss, suggesting the effectiveness of transferring DINOv2 features to BIM component retrieval. Although TripoSR-based generation is currently limited by data scarcity and input texture quality, it shows promising potential as a complementary pathway. The proposed framework offers a scalable and cost-effective solution for automated LOD3 reconstruction, contributing to high-fidelity digital twins in complex urban environments.