Benchmark
Registration residuals across 41 control networks
Cloud-to-cloud registration compared against total station control on 41 sites. Median residual 1.8mm at 20m, with the full per-site table and the six cases that failed tolerance.
Research index
Everything we assert about accuracy on this site traces back to one of the documents below. Where a claim has no published basis, we do not make it.
Publications
Method note
Photogrammetric scale drift in long corridor captures
Why unbraced corridor runs over 60m accumulate scale error, how much, and the two control placements that remove it.
Benchmark
Sensor comparison: phase-shift versus time-of-flight on wet surfaces
Reflective and wet surfaces degrade both sensor families, but not equally. Measured degradation curves for four instrument classes.
White paper
What a digital twin actually has to guarantee to be useful for maintenance
A twin that is geometrically pretty but has no provenance record is decoration. Proposes a minimum metadata set for asset-owner handover.
Method note
Decimation strategies that preserve dimensional intent
Aggressive mesh decimation quietly destroys the features engineers measure from. Comparison of five strategies against as-built tolerances.
Dataset
Open point cloud: nineteenth-century pump house, full control network
A complete capture released for teaching and software validation, including raw scans, control coordinates and the registration report.
How we publish
Failures are included
The Q3 benchmark covers 41 control networks. Six of them failed the tolerance we set before starting. All six are in the report with their residual tables, because a benchmark that only reports successes measures nothing except our willingness to discard data.
Instruments are named, vendors are not consulted
We name instrument classes and firmware versions in comparison work. No manufacturer reviews a report before publication and no manufacturer funds this programme.
Datasets before conclusions
Where licensing allows, the underlying dataset is released with the report so the analysis can be repeated. Client data is excluded unless the client has given written permission, in which case it is anonymised at site level.
Corrections
Two reports have been corrected since 2024. Corrections are published as amended versions with a change log at the front, and the superseded version stays downloadable.
Open datasets
Released for teaching, software validation and independent re-analysis under CC BY 4.0. Attribution is appreciated; permission is not required.
| Dataset | Contents | Size | Licence |
|---|---|---|---|
| Pump house, 1887 | Raw scans, control network, registration report | 4.2 GB | CC BY 4.0 |
| Corridor drift study | Photogrammetric imagery, 3 control configurations | 11.6 GB | CC BY 4.0 |
| Wet surface comparison | Paired captures, 4 instrument classes | 7.9 GB | CC BY 4.0 |
Datasets are distributed on request rather than by direct download, because the bandwidth cost of hosting them publicly is not something a lab of this size can absorb. Ask and we will send a transfer link the same week.