Why Your 3D Scanner Misses Its Accuracy Spec

You paid for a scanner rated at 0.05mm, scanned a machined bracket, and then measured the mesh against the part. The bore is close. The distance between two bores 400mm apart is out by three tenths of a millimetre. Nothing is broken, nobody lied, and the number on the box was true when it was measured — the 3D scanner accuracy spec is a statement about a specific measurement made under specific conditions, and almost none of those conditions were present on your bench.

This post contains affiliate links. If you purchase through these links, 3D Prototyping Hub may earn a small commission at no extra cost to you.

This post also contains Amazon affiliate links. As an Amazon Associate I earn from qualifying purchases.

This is worth an hour of your attention once, because the gap between the datasheet and your part is made of four or five named things, each with a number attached. Once you can name them you can decide which ones to fix, which to design around, and when the honest answer is that the job belongs on a different instrument.

The Datasheet Describes Three Different Things

Most confusion here is a vocabulary problem. Three figures get used interchangeably in marketing copy and mean entirely separate things.

Accuracy is the headline, and it is usually local. It is measured over a small patch, in one frame or a few, at the working distance the manufacturer specifies, against a calibrated artefact — a matte sphere or a sphere bar — under the kind of conditions optical systems are specified against in VDI/VDE 2634 Part 3. Your part is not that artefact and your workshop is not that room.

Volumetric accuracy is the figure that governs assemblies, and it is the one most consumer datasheets omit. Written as a fixed term plus a per-metre term, it says the error grows with the size of the thing you scanned. That per-metre term is where a part that "measured fine" turns into an assembly that does not fit.

Point distance is density, not position. It decides whether a 0.3mm engraved line arrives as geometry or as noise. A scanner can have excellent point spacing and ordinary accuracy, and at least one model on a major manufacturer's own store has shipped without any published accuracy figure at all — our 3D scanner buyer's guide flags that case, and the lesson generalises: if the figure is not published, ask for it in writing before you treat the device as an instrument.

One more pair worth separating. Repeatability is how closely five scans of the same part agree with each other. Accuracy is how closely they agree with the truth. A scanner that returns the same wrong answer five times in a row is repeatable and biased, and only one of those two problems is visible from inside the software.

Your Part Is Not the Calibration Artefact

Four conditions change between the lab and your bench, and each has a size.

The surface. Specs are measured on cooperative matte surfaces. Machined metal, chrome, gloss paint, matte black and anything translucent all break that, and the dangerous failure is not the missing data — it is the biased data, points that are present, plausible and shifted. A light pass of self-vanishing scanning spray puts the surface back into the condition the figure assumes. Why those surfaces defeat the sensor in the first place is covered in why scans come out noisy or full of holes.

The size. A handheld builds a model by aligning each frame to the ones before it, so alignment error compounds along the part. Local geometry stays crisp while the global shape bends — a 600mm extrusion comes back gently banana-shaped with every hole individually perfect. Positioning marker dots are the direct fix, because they give the software fixed references on the part rather than a chain of relative ones. Put them on the fixture and the surrounding faces, never on a feature you intend to measure.

The temperature. This one is routinely blamed on the scanner and belongs to the part. Aluminium expands about 23µm per metre per degree C, steel about 11 to 13. A 500mm aluminium part in a 30°C shop measures roughly 0.12mm longer than the same part at 20°C, which is where its drawing nominal lives — larger than the instrument's own error, and no amount of rescanning removes it. Let the part and the scanner equalise in the room, and write the temperature down next to the result.

The calibration. Calibration is a statement about the optics at a point in time: a focus ring that moved, a knock in transit, a different ambient temperature, and the geometry the software assumes no longer matches the geometry it has. A fixed rig on a tripod holds that geometry far better than a handheld does, which is most of what you are paying for in a standalone structured-light system like SHINING 3D's EinScan line — a set working distance and a calibrated working volume instead of two conditions held by your arm. Recalibrate on the manufacturer's schedule, and again after anything that counts as an event.

Keeping distance and overlap constant removes a fifth variable that nobody lists, and for anything bench-sized a motorised turntable does it better than a careful operator can.

Accuracy You Lose After the Capture

The mesh you measure is not the data the sensor produced, and three post-processing steps move the surface before you ever get to it.

Noise reduction and smoothing. Both work by moving measured points toward their neighbours. On a noisy organic capture that is a net improvement; on a machined flat it rounds the edge you were trying to measure, and on a bore it shrinks the radius slightly. Note the smoothing level you ran, because it belongs in any accuracy claim you make about the result.

Decimation. Sensible decimation introduces surface error well below the scanner's accuracy and is worth doing. Aggressive decimation to make a file manageable is a geometric edit, and it hits high-curvature features — fillets, chamfers, small bores — hardest, because that is where the triangles it is removing were doing work.

The alignment you compare with. This is the big one and it is a choice, not a setting. A best-fit alignment to CAD distributes deviation evenly across the part, which produces a reassuring rainbow map and conceals whether a feature is actually in position. A datum alignment constrains the scan the way the drawing constrains the part, and shows you the error the inspector would see. Comparing against a best fit and reporting it as conformance is the most common way a correct scan produces a wrong conclusion — a distinct problem from the geometry issues in why 3D scans don't work in CAD.

Measure Your Own Number in an Afternoon

The published figure belongs to the manufacturer. The figure that matters belongs to your room, your surfaces and your hands, and you can have it today.

1. Pick a known reference. A gauge block, a ring gauge, a precision sphere, or a machined feature you have measured carefully. Scan something whose dimensions you did not get from the same scanner.
2. Measure the physical part properly. One short feature around 10mm and one as long as your typical work, with digital calipers and a consistent technique. Record the temperature.
3. Scan it five times, breaking down and re-setting between scans, exactly as you would on a real job.
4. Measure the same two features in all five meshes, with the same smoothing and decimation you normally apply.
5. Read two numbers. The spread across the five scans is your repeatability. The difference between their average and the known value is your bias. The long feature exposes the per-metre behaviour the short one cannot.

Repeat it on the surfaces you actually scan — bare machined metal, sprayed metal, black plastic — and you will have measured the thing the datasheet could not: this instrument, in this room, on this material, in these hands. The way those numbers relate to a printed or machined tolerance is laid out in tolerances and accuracy in 3D printing.

Which Figure Should Decide the Purchase

Match the specification to the job rather than buying the biggest number available.

• Shape capture — props, figures, organic forms, printing a copy. Accuracy is close to irrelevant and point spacing is what you feel. Nearly any current handheld clears this bar.
• Fit-up against an existing object — a bracket for a car, a cover for a machine, a replacement for a part you still have. Local accuracy plus markers plus discipline about length. Revopoint's handheld line sits in this band, with its blue-light models specified tighter on small objects; read the per-model figures rather than the line's headline.
• Dimensional work against a drawing — inspection, first articles, a supplier dispute. This needs a published volumetric figure, verification against a traceable artefact, and a documented method. That is standalone structured-light territory or a service, and the reason is procedural as much as optical: you have to be able to show how the number was obtained.

When the Answer Is a Report, Not a Mesh

There is a clean line here. If what you need is geometry, buy a scanner and learn it. If what you need is a statement — this feature is within this tolerance, measured this way, at this temperature, on an instrument with current calibration — then you are buying metrology, and metrology is a process with paperwork rather than a device.

Shops set up for it hand you a deviation report aligned to your datums, not a mesh with a colour map. What to ask for and how those quotes are structured is in our guide to 3D scanning services for quality inspection, and the provider directory lists shops by location and process so you can compare two or three before deciding whether to buy the instrument at all.

A scanner's accuracy figure is a measurement of an instrument. Your part's accuracy is a measurement of a process — the instrument, the surface, the temperature, the registration and every click you made afterwards. The spec is the ceiling. Everything in this article is the distance between you and it.

Hero photo by Kelly No.715 via Wikimedia Commons (CC BY-SA 4.0).