Dataset and benchmark for as-built BIM reconstruction from real-world point cloud

As-built BIM reconstruction plays a significant role in urban renewal and building digitization but currently faces challenges of low efficiency. Scan-to-BIM aims to improve reconstruction efficiency but lacks domain-specific, large-scale datasets and accurate, multi-dimensional benchmark metrics. These deficiencies further impede the evaluation and training of scan-to-BIM methods. To address these challenges, this paper proposes BIMNet, an IFC-based large-scale point cloud to BIM dataset, and a set of metrics that reflect the quality and issues of reconstructed models from both geometric and topological perspectives. Experiments demonstrate that BIMNet enhances the evaluation and training of scan-to-BIM methods during the critical processes of reconstruction and segmentation. This research contributes to the data foundation and metric system for deep-learning based scan-to-BIM methods. In the future, BIMNet will not only facilitate the development of scan-to-BIM but also contribute to the advancement of smart cities and AI-driven technologies beyond scan-to-BIM.