Math

Challenge

A scanner's headline accuracy number hides a sum of mechanisms: sensor noise, calibration bias, registration drift, fusion artifacts. Improving any of them first requires instruments that can tell them apart against real ground truth.

Contribution

Built evaluation instruments using methods informed by ISO 20896-1:2019, and investigated the error budget over a 17-scan corpus against industrial ground truth. Comparing individual frames, estimated motion, and reconstructed surfaces helped distinguish errors already present in the measurements from those introduced downstream.

Characterized a ±6 µm periodic error in the calibration stage's leadscrew, which the standard sweep could not have resolved: a periodic signal can only be seen if you sample it more than twice per cycle — its Nyquist rate — and the sweep stepped more coarsely than that, folding the ripple into what read as noise.

A later investigation found an approximately −0.2% lateral scale error in all 17 scans. The calibration target's assumed feature spacing was a plausible explanation, but its true spacing had not been measured against a traceable standard, and a smooth bias in feature detection could produce the same signature. Differences between the recorded corpus and the development setup also prevented transferring a correction directly between them.

Outcome

Evaluation tools that made the remaining uncertainties more specific and testable. No calibration change compensating for the assumed feature spacing was shipped: the expected gain was small, and further measurements were needed to establish whether a correction would hold across development and field setups.