scan‑space

Measured accuracy

These numbers are not a claim, they are a test result. The solver is run against a synthetic capture of a flat whose true dimensions are known to the inch, with an ARCore-realistic noise and drift model. Re-run it with go run ./cmd/validate.

How to get an accurate scan

1. Calibrate scale first. Tap two points on one wall and type the real distance between them. Visual odometry has no absolute scale; one tape measurement removes a 1–3% error from every dimension in the result.
2. Sweep, do not spin. Depth comes from parallax between frames. Standing still and rotating gives the phone no baseline. Step sideways as you pan.
3. Hold at chest height, tilt slowly. The floor and ceiling fix level and storey height; both need to be seen at a decent angle, not edge on.
4. Pause in every corner. Corners are where the Manhattan constraint gets its evidence. Two seconds each is enough.
5. Walk through every doorway. Doors are found by rays that pass through a wall plane, so walking the opening is what carves it.
6. Aim at windows deliberately. Depth returns nothing through glass; that absence is the signal, but only where the phone actually looked.
7. Close the loop. Finish where you started. Drift is redistributed rather than accumulated, which is worth roughly 3× on final accuracy.
8. Flag mirrors. A mirror reads as a hole into a room that is not there. Tap it so the depth behind it is discarded.