Chapter 01 of 04
The brief
A scan already knows the shape of a building, but a buyer wants a floor plan: walls, rooms and doors, drawn to scale on one page.
Six generations of pipeline tried to get there, ending in a machine-learning model built on published architectural-reconstruction research. The version that ships is not that one.
A scan carries the building's true geometry, and almost nobody wants to look at a point cloud. What sells a property is the drawing every buyer already knows how to read: a plan, to scale, with the rooms named and the doors in the right places. Between those two things sits the actual problem, which is deciding what counts as a wall.
That question is much harder than it sounds, because a scan of a lived-in home is full of things that look like walls and are not. Wardrobes, curtains, kitchen islands and half-height partitions all read as vertical surfaces, and a plan that promotes any of them to architecture is wrong in a way a client will spot immediately.
Chapter 02 of 04
The problem
Turn a 3D scan of a real property into a floor plan a buyer can read: walls, rooms and doors, drawn to scale on a single page.
Chapter 03 of 04
The build
- Research
- Scored six approaches against each other on the same eighteen scans rather than assuming the newest was best. The comparison is the deliverable: without it the sophisticated pipeline would have shipped on the strength of being sophisticated.
- AI & Machine Learning
- Built the learned pipeline properly before judging it, including auto-labelling the training scans, two rounds of fine-tuning and a scoring overhaul, so its loss could not be written off as a half-hearted implementation.
- 3D & Spatial
- Diagnosed the loss rather than accepting it: the model learns from synthetic buildings and is asked to draw real scanned ones, and it consistently under-covers the true area, which costs it more than accurate room detection wins back.
- Data Pipelines
- Kept every generation runnable behind one orchestrator, so the comparison could be re-run on new scans and the method that ships stays whichever one actually scores best.
Build notes. How it came together.
Chapter 04 of 04
What it did
The plain methods won.
Scored on the same eighteen scans, the machine-learning pipeline won on two, still two after fine-tuning, and three only when paired with a plain method's building outline. So the plans published here were drawn by plain methods, and each title block names its own: the apartment by grid_v9, the villa by v11 footprint. Six is how many generations the comparison keeps side by side; grid_v9 is the source repo's ninth pipeline version and the third of those six.
The apartment plan published on this site was drawn by the ninth version, a plain grid method, and its own title block says so. Anyone can check: the label reads grid_v9. The villa beside it reads v11_footprint, an earlier plain method, and that is the point rather than an exception: both published plans were drawn by versions the learned pipeline was built to supersede. That is the honest state of this work, and shipping the version that scores best rather than the version that sounds best is the whole point of having built the scoreboard. The learned pipeline is not abandoned; it is waiting on being paired with something that fixes the area shortfall, which is a different piece of work from tuning it harder.
- 6
- Generations builtgeometry through to ML
- 2 / 18
- ML win rate3 as an ensemble
- 30 to 80%
- Area shortfallwhy it loses
Research origin.
[ Floor-plan extraction ]
Six pipeline generations, scored on the same eighteen scans, and the result: the simplest one won.


