The most common promise in reality capture is also the easiest one to test — and it is the one that fails. This essay brings together the studies that validated point clouds, BIM models and prefabrication checks against independent measurements: millimetric instruments, deliverables in centimetres, and standards that never promised millimetres. Companion to the scan-to-BIM radar and to the problems of LOD.
In one sentence: millimetre accuracy exists — on the instrument's datasheet and in well-controlled registrations — but the deliverable measured against an independent reference comes out in the regime of the centimetre, the accuracy standards themselves specify in bands of 15–50 mm, and no published validation study guarantees the millimetre as a confidence band across a whole building.
A German geodesist recounted the textbook case in September 2026: a joinery workshop was asking for a millimetre survey of an interior — not out of stated need, but because that was what the competition advertised.
Robin Miller (Ingenieurteam2 / Hexageo-Group) published the piece "Millimeter-accuracy" — Bullshit Bingo on LinkedIn after a client asked him for a millimetre as-built for prefabricated panels, justifying the request with a sentence worth quoting: many competitors advertise millimetre-accurate surveys using laser scanners. The firm's professional answer was that better than 5 mm would be difficult — and that even this figure should be treated with caution [C, Miller 2026 — an opinion piece by a practitioner, not a study; it is the trigger, not the evidence]. His diagnosis of the market the client had found:
Often, only a single instrument is used, without integration into a total-station control network, without independent check measurements and without reliable quality assurance of the resulting data. Miller (2026), LinkedIn · verbatim [C] · practitioner opinion, attributed
And the sentence that sums up the attitude: as long as the indicator is green, everything must be fine — the green indicator of the registration software as a substitute for metrological assurance. As the same author writes, a green status indicator is not proof of quality, because the software primarily evaluates the mathematical fit between the point clouds, not the geometric correctness of the result [C, same piece]. The structural point — and it is where his opinion meets the literature — is the confusion between two different things: anyone who turns a scanner specification into a blanket claim of a "millimetre-accurate survey" is confusing instrument specifications with demonstrable project accuracy [C, same piece].
Accuracy names at least three things in the same conversation — and the millimetre promise lives on mixing them up.
In the published numbers: the FGI reference cloud (Leica P40) has 3D position accuracy of 3 mm at a 50-m range and the registration of the stations reaches RMS error for registration … 2 mm and, on another dataset, 5.7 mm [C, Lehtola et al. 2017]. At the opposite end, the same study records what the manufacturer NavVis states for indoor SLAM work without control points: the absolute position accuracy … is 5 cm–50 cm after 10 min of walking [C, same study]. In between, the consumer LiDAR (Google Tango, predecessor of the phone sensors): ranging accuracy is 30 mm [C, same study]. Three orders of magnitude between the reference TLS and the stated SLAM — often sold with the same word.
And the bridge to the deliverable, in Miller's formulation: a registered point cloud is, first and foremost, a measurement product. It is not yet a finished geometric description of the building [C, Miller 2026]. The conversion to CAD/BIM adds point density, surface noise, the resolution of the orthophoto in mm/pixel, the viewing scale and — the least obvious and heaviest decision — which surface of the wall is the true one [C, same piece].
Eight validation studies, eight independent references, the same pattern: the instrument goes in millimetric, the deliverable comes out in centimetres — except when the geodetic method and the control exist, and even then the authors conclude in cm. One of them is Portuguese, presented at the 6th ptBIM Congress (FEUP, June 2026).
| Study | Independent reference | Measured result |
|---|---|---|
| Esfahani et al. 2021, Automation in Construction | Laser distance meter (1 mm up to 20 m); Faro Focus M70 scanner (±3 mm) | Modelled BIMs: mean absolute error 21.92–68 mm; σ up to 108.44 mm; in one scenario the error goes from 100 mm … to 490 mm; choosing the outermost or innermost point of the wall changes the result by 40 mm [C] |
| Bonduel et al. 2017, ISPRS | Leica P30 cloud; LOA30 specification (max. 15 mm at 95%) | Delivered model: 95% upper limit of the deviation = 17–19 mm — the top surface of the floor complies to LOA30, but the riser does not [C] |
| Skrzypczak et al. 2022, Building Research & Information | Trimble M3 total station (2″; 2 mm + 2 ppm) | n = 60: mean −1.6 mm, σ ±6.5 mm; the authors' conclusion: mean 3D BIM measurement accuracy ±6.8 mm and better than ±1 cm; under favourable conditions, up to ±0.5 mm [C] |
| Nascimento, Escórcio & Santos 2026, Atas do 6.º Congresso ptBIM | Manual measurement on site: laser distance meter + tape measure (expected error < 5 mm) | 4 case studies (18 reference dimensions each): overall mean error 6.3 cm; laser scanning 4.4 cm vs. photogrammetry 7.3 cm; BLK360: 4.3 cm (max. 35.1 cm); BLK2GO: 5.0 cm (max. 37.5 cm); up to 14.2 cm of mean in one photogrammetric test [C] |
| Lehtola et al. 2017, Remote Sensing | Leica P40 TLS (3 mm at 50 m) | TLS registration: RMS 2–5.7 mm; NavVis SLAM as stated by the manufacturer: 5–50 cm without control points; consumer LiDAR (Tango): 30 mm [C] |
| Tucci et al. 2018, Applied Sciences | Reference TLS in an internal/external test field | SLAM systems: σ of the signed distances 7 cm (path 1) and 5 cm (path 2); gross errors of as much as 79 cm downwards and 68 cm upwards in narrow passages; the best system: differences less than 5 cm [C] |
| Masiero et al. 2018, Applied Sciences | TLS-based reference model | Leica Pegasus backpack vs. TLS: RMS 8.2 cm (relative) and 16.1 cm (georeferenced); low-cost UWB-aided photogrammetry: 6.1 cm relative / 50.3 cm absolute; maximum of 193.4 cm, reduced to 79.3 cm by discarding one wing of the building [C] |
| Li et al. 2020, Applied Sciences | Stated manufacturing tolerances (±3/±5/±10 mm per axis) | 16 scans: registration < 5 mm, georeferencing < 5 mm, maximum measured deviation 5 mm on the Z axis, within the ±10 mm tolerance [C] |
The probability that a primary object is modeled within a ± 20 mm error range is equal to 49.2% with upper and lower confidence limits of 60% and 38.4%, respectively. Esfahani, Rausch, Sharif et al. (2021), Automation in Construction · scanner ±3 mm, 1 mm reference — and still one object in every two escapes the ±20 mm
It should be noted that equipment manufacturers point to errors of only a few millimetres (laser scanning) to a few centimetres (photogrammetry); these, however, result from tests under ideal conditions. Nascimento, Escórcio & Santos (2026), Atas do 6.º Congresso Português de BIM (ptBIM), pp. 140–149 · ptbim.org · read in full [C] — the same thesis, measured in Portugal and written in Portuguese
The cold reading of these numbers is not the laser is no good. It is that the accuracy of the deliverable is a property of the method (control network, targets, independent checks, modelling criterion, deviation report), not of the sensor — and most of the market contracts the sensor. And the Portuguese congress gets there by the same steps: the propagation of errors in the alignment of point clouds, the accumulation of inertial drift in SLAM algorithms, the dilation of positional uncertainty in the photogrammetric adjustment […] mean that the real errors are higher than those indicated by the manufacturers [C, Nascimento et al. 2026].
Two of the eight studies measure millimetres for real — and they state exactly what sets them apart from the others: control, short distance and stated tolerances.
Skrzypczak et al. (2022) is the optimistic extreme of the spectrum: complete geodetic method (total station as reference, TLS Faro), 60 comparisons, mean of −1.6 mm — and the conclusion of the authors themselves is prudent: the average accuracy of a 3D BIM measurement for inventory purposes will amount to ±6.8 mm, better than ±1 cm, reserving the ±0.5 mm for favourable conditions [C, Skrzypczak et al. 2022]. In other words: even the most favourable study in the sample concludes in centimetres — the mean may be millimetric, the spread is not. Li et al. (2020) show the other way to the millimetre: controlled environment, short distances, tolerance stated per axis and measurement against it — the maximum deviation measured 5 mm against the tolerance of ±10 mm [C, Li et al. 2020]. What these cases have in common is not the equipment: it is the verifiable method. And the opposite end — indoor SLAM without a control network — is where the 5–7 cm of σ and the gross errors of 79 cm live [C, Tucci et al. 2018].
The normative piece is missing: the American industry has already answered this, and the answer is not millimetric. The Level of Accuracy specification of the USIBD defines bands at the 95% confidence level: LOA20 = 15–50 mm, LOA30 = 5–15 mm, LOA40 = 1–5 mm, LOA50 = 0–1 mm [C, Bonduel et al. 2017, Tab. 2; USIBD guide ~ not read in full]. The Finnish COBIM suggests for BIM elements: corners 10 mm, surfaces 25 mm, irregular old structures 50 mm [C, same study]. The reference standards are written in centimetres — the millimetre promise is finer than what the industry itself considers contractually normal.
This collective delivers scan-to-BIM with the accuracy stated separately from the LOD: contractual LOA20, target LOA30 on most elements, non-conformances flagged (house statement). It is not tactical modesty: it is what the validation literature makes it possible to demonstrate. The millimetre is left for the scanner's datasheet and for the controlled cases with a geodetic method; the contract keeps the band an auditor can reproduce — and Miller's closer is the epitaph of the fine promise: even 100 millimetres are still … accurate to the millimetre [C, Miller 2026].
The instrument, yes (±2–3 mm of distance on the datasheet; a well-made TLS registration at 2–5.7 mm RMS [C]). The deliverable, no: against an independent reference, BIM models measure mean errors of 22–68 mm and limits of 17–19 mm at 95% [C]. The chain capture → registration → interpretation → modelling degrades into the regime of the cm.
They serve speed and completeness — layout, LOD 200, rough inventory, pre-diagnosis. The literature measures σ of 5–7 cm against TLS, with gross errors up to 79 cm in narrow passages, and the manufacturer states 5–50 cm without control. For prefabrication or fine coordination, it is not the tool.
The accuracy band 15–50 mm at 95% confidence on the LOA10–50 scale [C] — the normal contractual band of existing-conditions capture. LOA30 = 5–15 mm; LOA40 = 1–5 mm [C]. This collective contracts LOA20 with a LOA30 target.
It is achieved as a verified result: ±0.5 mm under favourable conditions with a geodetic method (Skrzypczak et al. 2022) and 5 mm of maximum deviation in a prefabrication QC with stated tolerances (Li et al. 2020). What no study measures is the millimetre as a confidence band across a whole building.
A band per element class (LOA scale), a verification method (control network, independent measurements) and a model↔cloud deviation report. Never the word millimetric without a band, a confidence level and a method — accuracy without verification is not a specification.
The quanto custa scan-to-BIM estimates price and lead time by area and says what goes into the delivery: LOD per purpose, accuracy stated in an LOA band, non-conformances flagged. It is the practical answer to the essay: contract what can be demonstrated.
Open quanto custa scan-to-BIM