Scan2BIM-Data: Automated IFC Generation for Text-to-BIM

The generative AI revolution (Text-to-Image, Text-to-3D) is rapidly approaching the architectural domain.
Text-to-BIM — generating parametric building models from prompts — is severely bottlenecked by data scarcity: there is no large-scale dataset pairing natural language with parametric IFC building structures.
Massive 3D point cloud datasets exist ( ScanNet , Matterport3D), but converting these unstructured geometric representations into structured, parametric architectural entities remains an open challenge in 3D computer vision.
This thesis addresses the generative architecture data bottleneck by automating the conversion of unstructured 3D point clouds into parametric IFC datasets.
Leveraging SOTA 3D semantic segmentation and geometric primitive fitting, the student builds a pipeline that extracts structural elements and exports them as IFC files, paving the way for open-source Text-to-BIM foundation models.

Semantic extraction: Use 3D vision models (e.g., PointGroup, Mask3D) to segment walls, floors, ceilings, and large furniture from point clouds.
Parametric fitting: Fit geometric primitives to segmented point clusters and translate them into standard IFC classes (e.g., IfcWall, IfcSlab).
Dataset generation: Generate a paired dataset of spatial descriptions and IFC files, and evaluate its viability for fine-tuning an LLM on simple Text-to-BIM tasks.

Python programming experience with PyTorch, NumPy, and Open3D
Deep learning knowledge, especially 3D point cloud networks (sparse convolutions, transformers)
Geometry processing and 3D data handling
Interest in generative models / LLM fine-tuning is a plus

Literature research on Scan-to-BIM and 3D semantic segmentation
Data preparation and preprocessing of large-scale indoor point cloud datasets
Implementation and training/evaluation of the segmentation and primitive-fitting pipeline
IFC export and construction of the paired language–IFC dataset

Bavishi, R., et al. [Meta Reality Labs] (2024). “SceneScript: Reconstructing Scenes with an Autoregressive Generative Model.” arXiv. — SOTA in language-driven structural reconstruction.
Dai, A., et al. (2017). “ScanNet: Richly-annotated 3D Reconstructions of Indoor Scenes.” CVPR. — Primary dataset source.
Schickert, M., et al. (2023). “Automated Scan-to-BIM: A review of current techniques and future directions.” Automation in Construction.

Ashwin Nedungadi