A BIM-Native Synthetic Data Generation Pipeline (BNSDG) for Instance-Level Floor Plan Segmentation

Abstract

Automating construction workflows such as drawing-to-BIM reconstruction, quantity take-off, and digital twin generation requires reliable extraction of structured information from engineering drawings. Deep learning offers a viable path toward this, but requires large-scale, instance-level annotated data, which is laborious to prepare manually and remains scarce in the AEC domain. Many existing BIM-based synthetic approaches rely on color-coded raster exports and post-processing, which can introduce misalignment and lose element metadata. This paper presents a BIM-native synthetic data generation (BNSDG) pipeline that produces training-ready image-annotation pairs directly from BIM geometry through parametric modeling and geometry-to-pixel mapping under the view crop-box transform, preserving BIM metadata. Geometric consistency is validated against a color-rendered baseline, and learning suitability is demonstrated with Mask R-CNN, achieving 62.0% bounding-box AP and 54.8% mask AP across seven categories. By automating dataset generation from BIM geometry, the pipeline reduces a data preparation bottleneck in AI-driven structural drawing interpretation.

Keywords: Building Information Modeling (BIM), Automated annotation, Synthetic Dataset Generation, instance segmentation, Floor plan interpretation