Accuracy, Precision & Error

The cleanest way to hold these two words apart is the classic target. Accuracy is how close your shots land to the bullseye — the true value. Precision is how close your shots land to *each other* — how repeatable they are, regardless of whether they are centred on the truth. Because these are independent, all four combinations exist. Shots tightly grouped on the bullseye are accurate and precise. Shots tightly grouped in one corner are precise but not accurate — consistent, repeatable, and consistently wrong. Shots scattered around the bullseye average out near the truth but are not precise. Shots scattered off-centre are neither.

The corner that causes the real damage is precise but not accurate, because it looks so convincing. A measuring system that repeats the same answer to a fraction of a millimetre feels authoritative; if that answer is systematically off, the authority is an illusion. This is exactly the point cloud that renders crisply, reads the same every time you query it, and is quietly biased because the instrument was miscalibrated or the registration drifted. The precision is real. The accuracy is absent. And nothing on screen warns you.

In reality capture both properties have concrete meaning. Precision shows up as the repeatability and the noise of the measurement: scan the same flat wall twice and see how closely the two clouds agree; look at how much the points fuzz around the true surface. Accuracy shows up as how closely the cloud matches reality as established by an independent, higher-trust measurement — a checked distance, a control point, a surveyed coordinate. You can improve precision by better instruments, more points and averaging; you improve accuracy by calibration, control and checking against truth. The practical discipline that follows is simple to state and easy to neglect: never infer accuracy from precision. A dataset that agrees with itself beautifully has told you only that it is precise. Whether it is *right* is a separate question that only an independent check can answer — which is why control points and ground-truth measurements, covered in Lesson 1.4, are non-negotiable for any capture whose correctness matters.

If accuracy and precision describe the *result*, systematic and random error describe the *causes*, and the two kinds behave so differently that telling them apart is essential. Random error is the unpredictable scatter in repeated measurements: sometimes a touch high, sometimes a touch low, with no fixed direction. It comes from sensor noise, tiny environmental wobble, the finite sharpness of any reading. Its signature is that it averages out — take many readings of the same thing and the random parts tend to cancel, so the mean converges on the truth and more data genuinely helps. Random error is what limits *precision*.

Systematic error, or bias, is entirely different: it pushes every reading in the *same* direction by a similar amount. A tape that has stretched reads long on every measurement; a scanner with a calibration offset places every point a few millimetres deep; a temperature effect biases a whole session. Its signature is the dangerous one: it does not average out. You can take a billion points, reducing random scatter to almost nothing, and the systematic part sits there untouched, shifting the entire dataset. Systematic error is what destroys *accuracy*, and because it is invisible within the data itself — everything is consistent, just consistently wrong — it can only be found by comparing against an independent reference that does not share the same bias.

This maps straight back onto the target. Random error is the spread of the shots; systematic error is the offset of their centre from the bullseye. More points and averaging tighten the spread but never move the centre. The reason this matters so much for reality capture is that the field's headline number — hundreds of millions of points — attacks only random error. It is wonderful for precision and seductive for confidence, and it does nothing whatsoever about bias. The habits that catch systematic error are different in kind: calibrate the instrument, tie the work to trustworthy control, check against an independent ground truth, and be suspicious of differences between methods. Whenever two independent measurements of the same thing disagree by a consistent amount, you have almost certainly found a systematic error, and that disagreement is far more informative than any amount of internal consistency.

A third confusion is specific to reality capture and just as costly: mistaking resolution for accuracy. Resolution is how finely the data samples the world — the point spacing or density, how close together the measured points sit, the smallest detail the capture can distinguish. Accuracy, as established, is how close those points are to the truth. They are independent, and the marketing of capture devices relentlessly blurs them, because point counts and resolution are easy to advertise and accuracy is hard to guarantee.

Two examples make the independence vivid. A very high-resolution cloud can be systematically inaccurate: pack points a millimetre apart on a wall while a calibration offset puts the whole surface five millimetres too far away, and you have a dense, detailed, wrong dataset. Conversely a sparse set of well-controlled survey points can be highly accurate: a handful of total-station positions, each checked and tied to control, may describe the true geometry far better than a dense cloud that was never checked. Detail and correctness do not imply each other. High resolution lets you *see* fine features and model them crisply; it says nothing about whether those features are in the right place.

There is a further subtlety worth internalising. Beyond a point, more resolution captures more noise and clutter, not more truth — you resolve the roughness of the plaster and the dust on the floor in exquisite detail, which is rarely what you needed and bloats the data (a theme of Module 5 on managing huge datasets). The professional question is never "how dense can I make it?" but "what resolution does this job actually need, and what accuracy must it meet?" — two separate specifications that you set independently. A heritage carving to be replicated needs high resolution *and* high accuracy; a floor plan for a fit-out needs modest resolution but dependable accuracy on the key dimensions. This is why the next lessons treat control, registration and checking as the spine of trustworthy capture: they are what supply accuracy, the property that resolution can never deliver on its own. When someone shows you a stunning, dense capture and calls it accurate, the correct response is a quiet question: accurate to what, checked against what, to what margin?

All of this converges on one mature idea: measurement uncertainty, the honest quantification of how much a measured value might differ from the truth. A professional measurement is never reported as a bare number; it is reported as a value plus an uncertainty — "3.200 m plus or minus 5 mm" — which says, in effect, the truth very probably lies within this band. Uncertainty is not an admission of incompetence; it is the opposite. Stating it is what distinguishes a rigorous measurement from a naive one, and refusing to state it is how false confidence takes hold.

Uncertainty has components that correspond to everything above. There is the random part, which you can estimate from repeatability — how much readings scatter. There is the systematic part, which you must estimate from calibration and from checks against independent references, since the data cannot reveal its own bias. And in reality capture there is a chain of contributions: the instrument's own measurement uncertainty, the effect of range, angle and surface (shiny, dark, wet and glassy surfaces misbehave), the registration that joins scans together, and any transformation into real-world coordinates. These combine, and errors can propagate and accumulate — a small per-station uncertainty can grow across a long chain of registered scans, which is precisely why the whole dataset can be less accurate than any single measurement within it.

The practical translation for a designer is the language of tolerance. Construction works to tolerances — the permitted deviation for a given element — and the uncertainty of your capture must be comfortably smaller than the tolerance of whatever you are going to do with it. Capturing a structure to plus or minus 20 mm is fine for a massing study and useless for manufacturing a component that must fit to 2 mm. The skill is to match the uncertainty you achieve to the uncertainty the task permits, and to say both out loud. This is also the firm edge of the course: any *stated, binding* accuracy — the number that goes on a deliverable and that others will rely on — is the province of a licensed surveyor and the equipment manufacturer's verified specifications, established by proper calibration and checking, not asserted from a screenshot. Every figure used in this course is illustrative of the principle; the real number, for a real job, is measured, checked and certified. Carry the habit: a measurement without a stated uncertainty is unfinished, and a capture you cannot check is a capture you cannot fully trust.