New tool automates building data labeling, cutting costs for AI training
Researchers have developed a method that automatically transfers semantic labels from building design models to 3D scans, eliminating months of manual annotation work. For construction firms and engineers deploying AI to analyze building performance and safety, this breakthrough could cut dataset preparation costs by half while accelerating AI model deployment in the field.
Originaltitel: Semantic annotation of 3D point clouds via label transfer from BIM models for AEC applications
<p>The lack of large, high-quality annotated datasets remains a fundamental bottleneck for 3D deep learning in architecture, engineering, and construction (AEC), where point-wise labeling of building-scale scans is costly and time-consuming. Many AEC workflows already produce Building Information Models (BIM) for design and documentation, providing an underutilized source of structured semantic information. We present a deterministic BIM-to-point-cloud label transfer pipeline that converts expert-authored BIM models into dense point-wise semantic annotations on real scans. Given a registered BIM-scan pair, object classes are propagated using FAISS-based nearest-neighbor search under explicit geometric tolerances, combined with vertical constraints and conflict filtering. The method enables scalable annotation without learning and serves as a tolerance-controlled measure of BIM-scan agreement. The approach is evaluated on the Stanford 3D Indoor Semantics Dataset (S3DIS) and two real-world buildings. For S3DIS, Areas 1 and 2 were reconstructed in BIM and used to relabel the original scans, enabling direct pointwise comparison without additional annotation. The pipeline is further demonstrated on a medieval castle (approximate to 8,000 m2) and a modern school building. On S3DIS Area 1, relabeling the same scan with a 25 mm tolerance yields high agreement with reference annotations (mean F1 = 88.54%) and 72.34% modeled-label coverage. Area 2 serves as a robustness study, with coverage remaining within five percentage points across tolerances. The method provides a scalable way to generate supervision directly on real scans when BIM-scan pairs are available and exposes unclassified regions as a QA signal for BIM incompleteness, misalignment, and geometric ambiguity. Results represent label agreement on the same scan rather than independent semantic accuracy.</p>