Floor Plan Extraction

Caution

Research project. For demonstration purposes only, not for productive use.

Exploring computer-vision extraction of structured floor plan data (rooms, room attributes, walls, openings, stairs, gross floor area) from heterogeneous plan archives for BBL's area management.

This repository brings together research, a target pipeline design and local pilot experiments. The pipeline is under development; the pilots document current findings and limitations.

Preview

• Live App: https://bbl-dres.github.io/plan-extract/

Documentation

• Motivation and goals — the area-management problem and project goals.
• Pipeline design — the target workflow, from uploaded plans to structured floor data and review.
• Literature review and state of the art — methods, evidence and remaining challenges.
• Research sources — the shared bibliography for the documentation and pilots.

Experiments

Pilot	Focus
Landgut Lohn · v1	Classical computer vision and a vision-language model on a first-floor plan
Graph pipeline · v2	A graph-based pipeline with a segmenter trained on style-randomised Swiss Dwellings renders

Each pilot includes its own scripts, findings and instructions. Plan data and derived overlays stay local.

Repository layout

Folder	Content
docs/	Motivation and goals, pipeline design, literature review and state of the art; solutions: open and closed; datasets: open, open building documentation and closed; research reports; research notes (e.g. reading DWG files)
docs/papers/	The shared source list, an index of papers and practitioner sources (papers.json), and scripts to download papers and convert them to Markdown
pilot/	Experiments on real plans: v1 (classical CV and a VLM on Landgut Lohn, 1. OG) and v2 (the graph pipeline with a segmenter trained on style-randomised Swiss Dwellings renders); scripts, findings and overlay viewers
data/	Training and evaluation data (local only)
assets/	README artwork and social preview

Data

Non-public training data everything derived from them (pilot/**/data/, data/), downloaded papers, their Markdown conversions and saved web pages stay local and are gitignored.

Licence

The code and documentation of this repository are under the MIT licence. Third-party material keeps its own licence: the curated plans and their images (data/curated/plans.json), the demo sheet (app/demo/README.md), the vendored three.js (app/js/vendor/three/LICENSE), and the datasets and models named in data/README.md and the pilot READMEs. The workflow app uses PyMuPDF (AGPL-3.0); the public instance at plan-extract.ch runs the code of this repository.