Accuracy

What we can prove, and what we will not claim.

Most scan-to-BIM suppliers quote a tolerance. We do not, and this page explains why that refusal is the honest position and what we publish instead: how much of your facade is found, and what evidence each element carries.

Completeness · one building, against a surveyor's IFC

How much of the facade comes back.

Scored element by element against the surveying firm's own IFC of the same building, modelled by hand from the same scans. The headline figure is how much of their model we found (recall); the line under it is how much of ours is theirs (precision).

Windows found
99%
and 94 % of the windows we drew are theirs (precision)
Doors found
100%
and every door we drew is theirs
Sills and bands found
99%
and every one we drew is theirs
Walls found
90%
and 94 % of the walls we drew are theirs
Wall surface covered
96%
of the reference surface; 100 % of ours lies on it
Elements carrying a measurement
98.6%
the rest name what was assumed, and why

One building, named: the estate's building A, against the surveying firm's own IFC of the same building. How much we found, recall, is the number we lean on because it survives what a tolerance does not: on the one pair we cannot register rigidly, how much we found holds, and every tolerance figure becomes a registration error instead.

Why no tolerance figure

A number we cannot defend is worth less than none.

The references disagree with each other more than we disagree with either. One cloud in our test set carries two independent third-party references. They agree with each other within 10 cm on every axis on 2 % of the 375 windows they share, at a median centre distance of 228 mm. Our own error against those labels sits inside that disagreement. No supplier can be certified against a reference that cannot agree with another reference over the same building.

Our certifiable base is one typology. The estate we can score against a survey is one kind of building. Scoring across it measures reproducibility, not generalisation, so any portfolio tolerance quoted from it would have to be withdrawn the first time you scanned something different.

The one pair that would test that cannot be registered. Its reference is three models merged in feet with independent offsets, and its two street faces are 0.785° out of square with the cloud's. Its errors are a registration term: they do not count against us, and they do not count for us either.

A claim that holds on the pair we cannot register is a claim we can defend anywhere.

The rule this page is written to

Per-element provenance

Every element says what the scan actually saw.

Each product carries Measured, Inferred and a NotMeasuredReason in its own property set. A wall footed on a ground line the scanner never reached says so, on the element, in the file you open.

That is the part you can check in your own model, without a survey — which no aggregate accuracy figure is, since a customer who had a survey to check us against would not need us.

It is audited rather than asserted: across 1,079 elements of four buildings, no element flagged as measured presented a face it had not measured, and the count stays at zero as the test is tightened from 5 cm to 40 cm. It does not claim every face of every element was measured, which is why the element also says which of its parts were.

On the elementWhat it tells you
Measuredthe cloud held the evidence for this dimension
Inferredit did not, and the value is a stated prior
NotMeasuredReasonwhat was missing, in words, on that element
MeasuredEdgeswhich parts of the element were measured: an opening's four edges, a wall's face and foot

Open the IFC in any viewer that shows property sets. Sort by Measured and you have your review list.

Say it first

What a one-sided scan cannot measure.

A terrestrial scan sees the face of a wall. It does not see through it. So we measure in the plane of the surface scanned, and anything behind that plane is a declared prior rather than a measurement.

  • Wall thickness is a prior, not a measurement, and it is labelled as one on the element.
  • We tested measuring it from the opening instead, on 443 windows across six datasets: it is worse than the stated prior every time, because what the scan can see there is the joinery's position in the wall, not the wall.
  • An opening behind a screen or a louvred stair tower returns no recess to fit; there the element keeps its proposed size and is marked unmeasured, with the reason.

This is true of everyone doing single-sided scan-to-BIM. We would rather you heard it from us than discovered it in a delivery.

If you need a number today

Per axis, on one named building.

The median absolute error over the 187 paired windows of the estate's building A, against the surveyor's IFC. Per axis rather than a single figure, because it tells you which axes to trust: the openings are where we say they are, and the depth is a prior.

AxisMedian errorWhat it means for you
Along the facade7 mmwhere an opening sits horizontally
Up the facade15 mmits height on the elevation
Width10 mmthe opening's size across
Height50 mmpart of this is the reference's own sill bearing, not ours
Depth36 mma prior, not a measurement: see above

One building, named, not averaged with anything. We will not turn this into a portfolio figure, and you should be wary of a supplier who does. The height figure is partly the reference's: its own window sits 21 mm above the bottom of its own opening, a property of that model rather than a fact about the building, and we are measuring how much of the 50 mm that accounts for.

What would let us claim a tolerance is not more algorithm work: it is one more surveyed building of a different kind, whose reference can be registered rigidly to its own cloud. That takes the evidence base from one typology to two. If you have such a pair and are willing to let us score against it, we would rather spend our time on that than on another decimal place.

Caveats

Read before you rely on it.

One estate is one typology

Every figure on this page comes from buildings of a single estate, captured and modelled by one firm. It is reproducibility. On buildings nobody had read before, on clouds four times thinner, we find 75 % of the windows — the honest generalisation figure, and the reason the completeness claim is stated per building rather than as a promise.

Two runs can differ slightly

Reading a facade is not perfectly repeatable: ten readings of one facade agreed on 87 of its 91 windows. We quote the run we shipped, not the best one we saw.

What is not modelled

Version 1 models facades: walls, windows, doors, vents, railings, terraces, copings, string courses and plinths. Mullions and screen rails have no detector. Interiors, roofs and terrain are not in this version, and no roof claim is made on this estate in either direction: three of its five buildings carry no roof geometry to score against.

The review exists because of all three. Nothing on this page is a reason to skip it.

Score it on your building.

If you have a reference model, the run can be scored against it and you get these figures for your own facade, not ours.