Why a 2 mm scanner does not give you a 2 mm process unit
Instrument accuracy, registration error and network accuracy are three different numbers. Manufacturers publish the first — Leica quotes 1.9 mm at 10 m for the RTC360, one sigma at 89% albedo. That figure describes one point from one setup. Across the setups needed to span a unit, registration and control decide your tolerance, and the datasheet stops being relevant.
Every scan project carries an error budget with four terms, and the manufacturer covers only one of them. The instrument contributes a range error and an angular error, both published, both stated at one sigma. The surface contributes another: reflectivity and incidence angle move the returned point, which is why Leica states its figures at 89% albedo and Trimble states range on a matte surface at normal incidence. Registration contributes a third, reported by the software as a bundle or cloud-to-cloud error. The control network contributes the fourth, and it is the only term that stops error from accumulating as setups chain across a unit. A scope of work that specifies only the scanner model has specified one term of four, which is why delivered data so often fails a tape check nobody agreed on in advance.
Source: USIBD Document C120, Level of Accuracy (LOA) Specification Guide, Version 2.0 (2016), current release Version 3.1 (2025); DIN 18710; ASTM E3125-17, Standard Test Method for Evaluating the Point-to-Point Distance Measurement Performance of Spherical Coordinate 3D Imaging Systems in the Medium Range; Leica Geosystems RTC360 datasheet (872750en, 07.21) and Cyclone REGISTER 360 error reporting documentation; FARO Focus Laser Scanner tech sheet (SFDC_04MKT_476, rev. 09/09/21); Trimble X7 datasheet (PN 022516-364A, 09/19).
| Error term | What it is | Published or measurable figure | How it is controlled | Accumulates across setups? |
|---|---|---|---|---|
| Instrument range error | Systematic and random error along the beam | Leica RTC360 1.0 mm + 10 ppm; FARO Focus S ±1 mm, defined as systematic error at around 10 m and 25 m; Trimble X7 2 mm | Factory calibration, FARO on-site compensation, Trimble's 25-second automatic range and angular calibration | No — the random part averages down, the systematic part is common to every setup |
| Instrument angular error | Error across the beam, opening into an arc as range grows | 18" (RTC360), 19" (Focus S), 21" (X7), all one sigma | Shorter shots and more setups; avoid documenting distant structure from a single far station | No, but it scales with standoff, so long shots become the weak links |
| Levelling and tilt | Error in the instrument's vertical axis at each station | Trimble X7 self-levelling accuracy <3" = 0.3 mm at 20 m; FARO dual-axis compensator 19" valid within ±2° | Level every setup and confirm the compensator is operating inside its stated inclination range | Yes if uncorrected — a tilt rotates the entire setup, not one point |
| Surface-induced error | Reflectivity, incidence angle, wet or polished surfaces displacing the returned point | FARO 10 m range noise 0.1 mm on 90% white against 0.9 mm on 2% black; Leica states its figures at 89% albedo, Trimble on matte surface at normal incidence | Approach surfaces square-on, shorten range on dark cladding, increase point density where returns are weak | No, but it is the term most often absent from a scope of work |
| Registration error | Residual disagreement between overlapping setups after alignment | Reported by Leica Cyclone REGISTER 360 as Bundle Error — the average of cloud-to-cloud and target errors — alongside Overlap and Strength | Targets, generous overlap, closed traverse loops, on-board IMU or visual tracking | Yes — this is the dominant term across a full unit |
| Network and control error | Error in the survey control the scan data is tied to | Set by the total station or GNSS survey and reported as network adjustment residuals | An independently observed and adjusted control network, established before scanning begins | No — control is precisely what stops accumulation |
| Represented error | Error introduced when the cloud is converted into linework or a model | USIBD C120 defines this as Represented Accuracy, held separate from Measured Accuracy | Specify an LOA band for the model as well as for the cloud, and check the model against the cloud | Adds on top of every term above it |
Three different numbers all get called accuracy
The word accuracy carries three distinct meanings on a scanning project, and conflating them is the root of nearly every dispute at handover. Instrument accuracy is a property of the hardware and describes one point measured from one station. Registration accuracy describes how well separate setups agree with one another after alignment. Network accuracy describes how well a point at one end of a unit relates to a point at the far end, and it is the only one that governs whether a fabricated spool arrives and fits.
USIBD's C120 specification separates these formally. Absolute Accuracy is defined there as a standard deviation related to a given reference frame such as a building or state coordinate system — a global figure over the whole object, and one that is independent of distance or the number of setups. Relative Accuracy is defined as a standard deviation related not to a superior datum but within an object's region or within one or more setups. The specification notes plainly that accuracy dilutes over distance through the concatenation of multiple setups.
That last clause is the whole subject in one sentence. A vendor can be truthful about a 2 mm scanner and still deliver a unit dataset where opposite ends disagree by far more, because concatenation is doing exactly what the standard says it does. The remedy is never a better scanner. It is control, and it must be specified in advance alongside the deliverable rung you are buying.
What the manufacturer's figure covers, and the conditions printed beneath it
Scanner datasheets are honest documents that get read carelessly. Leica publishes 3D point accuracy for the RTC360 of 1.9 mm at 10 m, 2.9 mm at 20 m and 5.3 mm at 40 m, built from a range accuracy of 1.0 mm + 10 ppm and an angular accuracy of 18 arcseconds. Beneath the table sits the qualifier that matters most: all accuracy specifications are stated at a level of confidence of 68 percent according to JCGM100:2008, and the figures apply at 89 percent albedo.
FARO states 3D point accuracy of 2 mm at 10 m and 3.5 mm at 25 m for the Focus S, with a ranging error of ±1 mm defined as a systematic measurement error at around 10 m and 25 m, and instructs that for distances larger than 25 m a further 0.1 mm per metre of uncertainty is added. Its footnote states that all accuracy specifications are one sigma, after warm-up and within the operating temperature range. Trimble publishes 2.4 mm at 10 m, 3.5 mm at 20 m and 6.0 mm at 40 m for the X7, at one sigma, valid when the instrument is levelled within ±5 degrees.
Three consequences follow. Accuracy degrades with range in every case, because the angular term opens into an arc. Every figure is one sigma, so restating it at the 95 percent confidence that engineering specifications use doubles it. And every figure applies to a warm, level instrument shooting a bright surface square-on. The datasheet describes laboratory conditions honestly; a process unit supplies none of them.
The surface is part of the measuring system
A scanner and the thing it points at form one measurement system, and the target contributes error the instrument cannot correct. FARO publishes the effect openly: range noise at 10 m of 0.1 mm against a 90 percent reflectivity white target and 0.9 mm against a 2 percent black target for the Focus S Plus, rising to 1.6 mm at 25 m on black. Maximum range collapses in parallel, from 0.6–350 m at 90 percent reflectivity to 0.6–50 m at 2 percent, stated for a Lambertian scatterer.
Incidence angle compounds reflectivity. A beam striking a pipe wall at a grazing angle spreads its footprint into an ellipse across a curved surface, and the returned point is computed from the centroid of an energy distribution that is no longer symmetric. Every one of these effects is largest on exactly the geometry that matters: the sides of vessels, the far faces of large-bore lines, and any surface documented from a single distant station because closer access was denied.
Industrial surfaces are systematically hostile. Weathered aluminium jacketing is specular rather than diffuse. Coke-covered and soot-blackened steel sits at the low end of the reflectivity scale. Wet surfaces after washdown scatter unpredictably. Damp or dusty air attenuates the return. None of this appears in a scope of work that names only a scanner model, and all of it degrades the same geometry you later want for mapping damage mechanisms across an asset.
How registration error is measured, and what a good report shows
Registration error is the disagreement between overlapping setups after alignment, and it is the term buyers most rarely ask to see. Leica's Cyclone REGISTER 360 defines Bundle Error as the overall indicator of the link quality forming the bundle, calculated as the average of cloud-to-cloud and target errors. Cloud-to-Cloud error is described as the average of all cloud-to-cloud errors encountered in the links, a reliable indicator directly reflecting the effectiveness of visual alignment. Target error is the average distance between pairs of connected targets in an individual setup.
Two supporting metrics carry as much information as the error figures themselves. Overlap is the average percentage of overlapping area in links, indicating how much two clouds share. Strength describes how well data is translated and represented across the X, Y and Z axes. A link with low error and low overlap is not a good link — it is a link with too little shared geometry for its error figure to mean anything. Trimble's Registration Assist reports project and station average error, overlap and consistency in the same spirit.
Demand the report as a deliverable and read three things. Whether every setup is connected by more than one link, since a single-link station has no redundancy and no way of revealing its own error. Whether overlap percentages support the reported errors. And whether the chain closes on control or simply runs from one end of the site to the other. Leica's own documentation pairs the numerical metrics with visual verification through slice tools, and that combination is the correct standard of proof.
How error accumulates when setups chain across a unit
Registration errors between successive setups are largely independent, which means they combine as the root sum of squares rather than adding directly. A chain of setups running from a unit's north boundary to its south accumulates disagreement steadily along its length, and the two ends are the worst-related pair of points in the entire dataset. Nothing about the instrument changes along that walk. The relationship between distant points does, and it degrades monotonically.
This is why a scan can pass every local check and fail the only one that matters. Measure a nozzle-to-nozzle distance inside one vessel and it agrees with the tape. Measure between a pipe rack support at one end of the unit and a foundation bolt at the other, and the disagreement is far larger than any datasheet suggests. USIBD anticipates this directly, noting that a relative standard deviation is used when higher demands arise for particular parts of a building, knowing that accuracy dilutes over distance through concatenation of multiple setups.
The practical consequence is that unit-scale tolerance must be specified as an absolute requirement against a defined reference frame, not as a relative one. It also means the question "what accuracy will I get" has two answers on the same project — a tight one within any single asset, and a looser one across the site — and both belong in the scope. Getting this right is what makes thickness data trended against real geometry reliable years later.
Control networks and targets are what break the chain
A control network is a set of physical points whose coordinates are established independently of the scanning, usually by total station traverse or GNSS observation, then adjusted and reported with residuals. Scans register to those points rather than to each other. The chain of accumulation is broken because every setup relates to the same external frame, and the network's own accuracy becomes the ceiling for the whole dataset instead of the sum of every link before it.
This is the single highest-value decision in a unit-scale scope, and it is routinely omitted because on-board registration works so convincingly at small scale. Modern instruments have made local alignment nearly effortless — Leica's visual inertial system tracks scanner movement between setups in real time, and Trimble's X7 performs full automatic calibration of range and angular systems in 25 seconds without user interaction or targets. Neither provides an external datum. They make a well-formed chain, not an anchored one.
Targets remain worth their field time on large jobs for a second reason: they are checkable. A surveyed target gives an independent coordinate that the registered cloud can be tested against, producing evidence rather than a vendor assurance. Specify the control network, its observation method, its adjustment report and the target layout in the scope, with the same discipline applied to writing a technical procedure that has to survive audit.
One vessel versus a whole unit: two different tolerances
A single vessel scanned from a handful of surrounding setups sits close to instrument specification. Ranges are short, so the angular term stays small. Links are few and mutually redundant, so registration error stays low and is easily checked. Surfaces are approached square-on because the scanner can be positioned freely around the object. This is the regime where 2 to 3 mm is a realistic expectation, and where deviation analysis against a nominal cylinder produces trustworthy out-of-roundness results.
A full process unit is governed by entirely different physics. Ranges lengthen because access is restricted. Setups multiply because congestion demands them. Links chain because a straight line of sight rarely exists between distant areas. Surfaces are shot at whatever angle the walkway permits. The achievable tolerance across such a dataset is set by the control survey, and expecting the scanner specification to describe it is a category error.
Write both into the scope explicitly. A clause reading "LOA30 relative within any individual asset, LOA20 absolute across the unit, both at the 95 percent confidence level stated in USIBD C120" is precise, achievable and verifiable. A clause reading "scanning to 2 mm accuracy" is none of those things, and it will be interpreted by the vendor in the way most favourable to the vendor — reasonably, because it is the only reading the sentence supports.
Writing an accuracy clause that survives delivery
USIBD C120 supplies the vocabulary. Its Level of Accuracy bands, specified at the 95 percent confidence level, run from LOA10 with a lower range of 5 cm, through LOA20 spanning 5 cm to 15 mm, LOA30 spanning 15 mm to 5 mm, LOA40 spanning 5 mm to 1 mm, to LOA50 spanning 1 mm to zero. The framework derives from DIN 18710, which prescribes five accuracy levels in terms of standard deviation, and it notes that any LOA higher than the one specified is acceptable because it exceeds the requirement.
The specification's most useful contribution to a scope is its split between Measured Accuracy and Represented Accuracy. Measured Accuracy is the standard deviation achieved by the final measurements, however acquired, affected by the sensor, by how measurements are joined into a common coordinate system, and by the object itself in terms of incidence angle, reflectivity, colour and roughness. Represented Accuracy is the standard deviation achieved once that data is processed into another form such as linework or a model. Error is always introduced in that conversion, so specify both.
Add the verification method and the pass criteria before mobilisation, never after delivery. ASTM E3125-17 provides the formal instrument-level analogue: it establishes requirements and test procedures for evaluating derived-point to derived-point distance measurement performance throughout the work volume of medium-range 3D imaging systems, from 2 m to 150 m, by comparing measured distance errors against manufacturer-specified maximum permissible errors. That structure — a declared figure, a defined test, an agreed pass mark — is what an enforceable clause looks like.
Verifying the data on the day it arrives
Three checks catch nearly everything, and all three are cheap. First, independent check distances: survey a set of distances between identifiable features at opposite ends of the site with a total station, then extract the same distances from the delivered cloud and compare. This tests network accuracy directly, which no software report does, and it is the only check that reveals accumulated chain error.
Second, read the registration report against the geometry rather than accepting its headline number. Confirm every setup carries more than one link, that overlap percentages are consistent with the reported cloud-to-cloud errors, and that the strength metric is balanced across all three axes. A dataset with excellent error figures and single-link stations has no redundancy, and no redundancy means no evidence.
Third, check coverage before checking accuracy, because missing data cannot be repaired at any price once the crew demobilises. Walk the coverage map against the asset list, confirm which elevations were captured, and get the inaccessible areas listed in writing. Atlantis writes these clauses, runs the capture, and delivers the verification evidence alongside the data. Affordable. Accessible. Fully customizable. Send an asset list and a tolerance requirement through contact for a quote or a demonstration.
What is the difference between instrument accuracy and network accuracy?
Instrument accuracy describes one measured point relative to the scanner head — the figure on the datasheet. Network accuracy describes how well a point at one end of the unit relates to a point at the other, hundreds of metres and many setups away. The second is always worse than the first, and it is the number that governs whether a tie-in fits.
How many setups before registration error dominates?
It dominates as soon as setups are chained rather than tied to control. Independent registration errors between successive setups combine as the root sum of squares, so the disagreement between the first and last station grows steadily along an unconstrained chain. Closing loops and observing control targets bound that growth; adding setups without either lets it run.
Do I still need targets if my scanner registers itself?
On-board tracking such as Leica's visual inertial system or Trimble's IMU-based Registration Assist handles local alignment well and removes most field target work over short chains. Across a full unit, surveyed targets or control anchor the whole network to something external. Read the registration report: bundle and cloud-to-cloud errors climbing along a chain are the signal that control is required.
How do I write an accuracy requirement into a scope of work?
State four things. The USIBD LOA band required, separately for Measured Accuracy and Represented Accuracy. The confidence level, since LOA is written at 95 percent and datasheets at one sigma. Whether the requirement is absolute or relative, and over what extent. And the verification method to be applied at handover, with pass criteria agreed before mobilisation.
How is scan accuracy verified when the data arrives?
Independent check measurements, agreed in advance. Surveyed check distances between features at opposite ends of the unit, compared against the same distances extracted from the cloud, test network accuracy directly. ASTM E3125-17 formalises the equivalent instrument-level test, evaluating derived-point to derived-point distance errors against manufacturer-specified maximum permissible errors across the work volume.
What tolerance is achievable across a whole unit versus one vessel?
They are different problems. One vessel scanned from a few setups sits close to the instrument specification, because few registration links exist. A full unit spanning hundreds of metres is governed by the control network, and its achievable tolerance is set by the survey that established that control rather than by the scanner. Specify each extent separately.