Interactive Demo

Step the Pipeline

A floor plan raster is only pixels. Architectural software needs objects — walls that meet at junctions, doors hosted in those walls, rooms you can walk between. Pick a plan and step through the six stages that recover them.

Stage 1 / 6

Input

Full size ↗

What you are looking at

Stage images and the final vector are real artifacts written by the pipeline, replayed here as static files. Nothing runs in your browser: the wall-graph model is a 397 MB Deformable-DETR-style decoder and the geometry stage is several thousand lines of Python. The three plans shown are training-split samples, so treat them as a walkthrough of the method rather than as held-out performance.

Attribution

Source floor plans come from CubiCasa5K (high_quality_architectural subset) — Kalervo, Ylioinas, Häikiö, Karhu and Kannala, “CubiCasa5K: A Dataset and an Improved Multi-Task Model for Floorplan Image Analysis,” SCIA 2019. github.com/CubiCasa/CubiCasa5k

CubiCasa5K is licensed CC BY-NC 4.0. The inputs on this page and the pipeline output derived from them are adaptations of CubiCasa5K and remain CC BY-NC 4.0 — shareable with attribution, not for commercial use. They are not covered by the source repository’s GPL grant.

Code

Pipeline and this page: daegeun-kim/neural_floorplan, licensed GPL-3.0-only. Built on Raster-to-Graph (GPL-3.0) and nvidia/mit-b0, each under its own terms — see NOTICE in the repository.