Semantic annotation of 3D point clouds via label transfer from BIM models for AEC applications

The lack of large, annotated datasets is a major bottleneck for 3D deep learning in the built environment. Unlike 2D images, 3D point clouds are costly and slow to label, especially in architecture, engineering, and construction (AEC). We present a pipeline that transfers semantic labels from expert-authored Building Information Modeling (BIM) data to point clouds. After aligning manually created BIM models with laser scans, object-level classes are propagated to individual points using coordinate alignment, FAISS-based nearest-neighbor search, and proximity thresholds, enabling efficient annotation of millions of points. We demonstrate the approach on the Stanford 3D Indoor Semantics Dataset (S3DIS). A subset (Area 1) was recreated in Autodesk Revit, and the resulting BIM was used to label the original S3DIS point clouds, enabling direct comparison with learning-based methods on identical inputs, without manual annotation. We further apply the pipeline to a scan of the medieval Kalmar Castle in southern Sweden (≈8,000 m2) and to a modern school building. Selected annotated point clouds are shared and can be used to train AI models or as weak supervision to reduce labeling effort. The method offers a scalable, domain-aligned alternative for generating labeled 3D datasets from existing BIM assets. It leverages expert decisions embedded in design workflows, produces point-level labels within user-defined tolerances, and is robust across historical and contemporary buildings. By converting produced BIM into supervision, the pipeline helps address the data bottleneck in AEC and supports development and evaluation of segmentation methods without large manual annotation campaigns.