accuracy benchmarks
Measured, with the method next to every number.
These results come from real event galleries, scored against AWS Rekognition as a reference and against known impostors. The write-ups are internal and not yet published. Where we have not measured something, this page says so instead of quoting someone else's number.
Last reviewed 2026-10-27
Current default: threshold 0.42
Current default- Measured
- 2026-07-08
- Evidence
- Internal and unpublished. The full write-up for this study is not available, so treat these figures as indicative until it is published.
- Operating point
- Cosine similarity 0.42, the default match threshold. Since 2026-08-09 a trial has served 0.40 to a 10% sample of galleries of up to 10,000 faces; everything else uses 0.42. Compare returns the threshold it applied in each response.
- Data
- Real event galleries from production: 618 genuine selfie searches across galleries from about 130 to about 990,000 faces.
- Reference
- For each selfie, the photos AWS Rekognition returned for the same selfie. Rekognition is a reference, not ground truth.
Mean recall vs Rekognition
74.7%
618 genuine selfies. 81.0% at threshold 0.36.
Zero-result rate
4.0%
Genuine selfies that found nothing. 1.9% at threshold 0.36.
Method
- Recall vs Rekognition: the share of Rekognition's photos for a selfie that we also returned, averaged over selfies.
- Zero-result rate: the share of genuine selfies that got no photos back at all.
- The same searches were scored at several thresholds, so the trade-off is measured, not modelled.
Caveats
- Recall here is agreement with Rekognition, not ground-truth recall. Rekognition misses faces too, and a photo only we found counts as neither a hit nor a miss.
- We raised the default from 0.36 to 0.42 because impostor testing showed false identifications of strangers fall sharply at 0.42, concentrated in galleries above about 2,000 faces. We do not publish a single false-identification rate for this operating point yet: our internal measurements do not reconcile closely enough to quote one figure. The 1:1 false-match rate in the study below is the published precision-side evidence.
- Recall falls as galleries grow (see the size table below). A single average hides that.
Agreement with Rekognition: 50 events
Historical operating point- Measured
- 2026-06-04
- Evidence
- Internal write-up, not published. The figures below are quoted from it as written.
- Operating point
- Cosine similarity 0.38 with the quality gates in use at the time. This is NOT today's default, and recall at today's stricter 0.42 is lower (see above).
- Data
- 50 production event galleries (weddings, corporate events, races, convocations, schools), about 4.06 million indexed faces and 22,070 registered selfies.
- Reference
- For each selfie, the photos AWS Rekognition returned at a 90% face-match threshold, the production setting at the time. Rekognition is a reference, not ground truth. The impostor false-match rate below uses account identity instead, with no Rekognition involvement.
F1 vs Rekognition (pooled)
0.824
Precision 0.798, recall 0.851.
Median per-event F1
0.883
Half the events scored above this, half below. Per-event results vary widely; the pooled figure above is the honest headline.
Impostor false-match rate (1:1)
≤0.13%
Selfies from different accounts, at threshold 0.38. On the same pairs, today's stricter 0.42 can only accept the same number or fewer.
Selfie face-detection coverage
96.9–100%
Share of selfies Rekognition searched that we also found a face in, across the five events reviewed in depth.
Method
- Photo-level matching: a photo counts as a match if any face in it clears the threshold.
- One global threshold for all 50 events, not a per-event optimum (per-event tuning would have flattered the result by overfitting).
- Precision, recall and F1 pooled over every selfie-photo decision (micro), plus the median of per-event F1.
- False-match rate: 1:1 comparisons of selfies registered by different accounts, so the result does not depend on Rekognition.
- Claims that did not survive internal review were retracted in the write-up, not quietly dropped.
Caveats
- F1, precision and recall here measure agreement with Rekognition, not absolute accuracy. They do not show that we match Rekognition, or that either system is right.
- Measured precision is a floor. Rekognition's detector misses faces ours finds, so some of our “false positives” are photos Rekognition never indexed. A follow-up test on five events, ranking each disputed match against every registered attendee, put true precision at roughly 0.88–0.91, not higher.
- Most remaining errors are relatives and lookalikes, concentrated at weddings and family events, not random strangers.
- Some accounts register several people (a family on one email). The 1:1 false-match test treats different accounts as different people, so a few of its “impostor” pairs are probably the same person.
Recall by gallery size
Bigger galleries mean more chances for a lookalike, so recall falls as size grows. Same study as the current default, split by how many faces the gallery holds, at threshold 0.42. False identifications also concentrate in the largest galleries; we do not publish a per-size rate because the impostor sample per size is too small to quote.
| Gallery | Faces | Recall vs Rekognition |
|---|---|---|
| Small | 130–492 | 88.9% |
| Medium | 542–1,986 | 77.5% |
| Large | 2,087–9,990 | 77.2% |
| Very large | 10,000–990,000 | 68.7% |
Public datasets: not run yet
The studies above use private galleries, so you can't rerun them. Next is a public-dataset run through the same public API you use, with the raw scores published so anyone can check it. It has not been run, so there is no result here yet.
| Benchmark | Protocol | Metrics | Result |
|---|---|---|---|
| LFW verification | Standard 10-fold, 6,000 pairs, unaligned images sent to Compare | Accuracy (mean ± std), TAR at FAR 0.1% and 1%, ROC | Not run yet |
| LFW 1:N search | 3,749-face gallery; 7,484 mated and 2,000 non-mated probes through Index and Search | Rank-1/5/10, FPIR and TPIR across thresholds | Not run yet |
Thresholds
Every match carries a cosine similarity score from 0 to 1. The default match threshold is 0.42, chosen to keep false identifications of strangers low in event photography, at the cost of some recall. Since 2026-08-09 a trial has served 0.40 to a 10% sample of galleries of up to 10,000 faces. That trade-off is yours to change: pass threshold on search and choose accept, review and reject bands from your own labelled pairs. Numbers don't carry across vendors, or even across use cases.
Limitations
- None of the numbers on this page can be rerun by you. They come from private production galleries, which is exactly why a public-dataset run is next.
- Rekognition is the reference, not the truth. Agreement with it is not the same as being right, in either direction.
- Every result is from event photography: phone selfies searched against professional event photos. Identity verification, CCTV and access control behave differently.
- Bad selfies dominate the failures we have seen: photos of a screen, heavy coloured lighting, a face that is barely visible. Validate selfies at capture and ask for a retake.
- Some people cannot be found at any threshold, because the selfie does not resemble the photos closely enough. Threshold tuning is not a universal fix.
- Results are tied to the model and thresholds in use when measured. They are re-measured when either changes, and the date is on each study.
Bias and fairness
- Children's faces are a known weak spot. In the 50-event study (threshold 0.38, recall measured against Rekognition), median recall at a primary-school event was 0.78 even when we kept only high-quality selfies, against 0.91–0.93 at adult events. Better selfies did not close the gap, so it is largely a limit of the model, not of the inputs.
- We have not yet measured accuracy by skin tone, gender or age group, and we will not claim parity across groups until we have. Face-recognition models have documented accuracy gaps between demographic groups, and you should assume ours may too.
- If your population is mostly children, or differs from adult event guests, test on your own consented data before relying on any number here, and keep a human confirmation step for borderline matches.
- Public benchmarks such as LFW skew heavily toward adult public figures, so even our planned public run will not answer the fairness question. A demographically balanced evaluation is a separate piece of work.
How we handle consent, isolation and deletion is on Trust & responsible use.
Test it on your own photos
The only benchmark that settles it is yours. Score a few hundred labelled pairs from your own data and pick your thresholds from that.