A raw scan of a room is not a floor plan. It is a few million points in space, some of which are walls, most of which are a sofa, a bin, a cat, and the person holding the phone.
Getting from one to the other is a sequence of decisions about what to throw away.
Step one: find the planes
The first job is separating structure from stuff. Large, flat, vertical clusters are wall candidates. Large, flat, horizontal clusters near the bottom of the scan are floor. Near the top, ceiling. Everything else is provisionally furniture and gets set aside.
This is where most naive approaches fall over: a tall bookcase flush against a wall looks exactly like a wall. The tell is thickness and continuity, real walls run the full height and connect to other walls at their ends.
Step two: regularise
Real walls are almost never perfectly straight or perfectly square, but they are meant to be. A plan that faithfully reproduces a two-degree wobble is less useful than one that snaps it.
We snap to dominant axes, square off corners within a tolerance, and close small gaps, while keeping a record of how much we moved things. If a wall was genuinely three degrees off, that is worth knowing, not hiding.
Fidelity and usefulness pull in opposite directions here. The right answer is to be honest about the correction rather than pretending it didn't happen.
Step three: openings
Doors and windows are found as holes in otherwise continuous wall planes, then classified by their height above the floor. A gap starting at floor level is a door. One starting at roughly waist height is a window. A gap running the full width with no top is usually the scan simply not reaching, and gets flagged rather than guessed.
Step four: vectorise and export
The regularised outline becomes polylines, openings become parametric symbols, and dimensions get attached to every wall run. From there it exports as PDF for humans, SVG for design tools, and DWG for anyone who has to build from it.
The failure modes worth knowing
- Mirrors. Reflections create phantom rooms. We detect and discard them, mostly.
- Glass. Depth sensors struggle with it. Large glazed walls sometimes need a manual nudge.
- Very dark rooms. Less light, noisier depth, softer edges.
None of these are exotic; they are just the parts where a human eye still beats the pipeline. The plan tells you where it was unsure, which is the next best thing to being right.
Why a walkthrough, not a single photo
A single photo, the input a [redesign](/blog/how-does-ai-interior-design-work) needs, gives you one angle, one moment, one perspective to work from. A floor plan needs to reconcile multiple rooms against each other: where the kitchen wall actually sits relative to the hallway, whether two rooms share a wall or just look like they might from separate photos. That requires continuity a single frame can't provide.
A phone walkthrough, walking through the space with the camera recording, the way you'd casually film a room for a friend, gives the model overlapping frames from slightly different angles and positions. That overlap is what lets it triangulate distances and reconcile one room's walls against the next room's, the same underlying principle behind photogrammetry generally, just applied to something people can do in under five minutes with a device they already own instead of specialist scanning hardware.
What "four minutes" is actually doing
The number isn't arbitrary, it's roughly the amount of walkthrough footage needed to give the reconstruction enough overlapping angles on every wall to resolve confidently, for a typical UK flat or small house. A single large open-plan room needs less; a home with many small rooms and doorways needs more, because each additional wall is another surface that needs to be seen from at least two angles to place accurately.
Where this is heading
The end goal isn't a pretty 3D model, it's a floor plan precise enough to trust with a furniture-fit decision, the exact gap [described in our piece on why room measurement is still hard](/blog/why-measuring-a-room-is-still-hard). That's a higher bar than "looks roughly right," and it's why this is still labelled coming soon rather than shipped: the difference between "close enough to visualise" and "close enough to buy against" is the whole remaining distance to travel.
