The results demonstrate competitive accuracy across multiple operational scenarios. Key performance metrics below are reported at a false match rate (FMR) of 0.00001.
5.45%
Best case — Mugshot-to-Mugshot
7.4–7.5%
Operational — aging & visa-to-border
37.9%+
Hardest cases — extreme angle & kiosk
FNMR at FMR = 0.00001 — lower is better
The algorithm performs strongly in constrained, cooperative environments — the scenarios most relevant to government ID verification, border control, and law enforcement applications. The Mugshot-to-Mugshot FNMR of 5.45% at a 1:100,000 false match rate places facehawk-001 within striking distance of established providers.
Large-scale identification (1:N) is where demographic bias typically shows up. NIST's official results score facehawk-001 at a GINI coefficient of 0.27 — among the fairest algorithms tested — with an MXOG (max-over-genotype) variation of 2.43, indicating minimal performance swing across demographic groups.
Industry average sits between 0.6 and 0.8 — lower is fairer.
Minimal false non-match rate variation across demographic groups.
False non-match rate by region and gender:
Consistent performance across all geographic regions and genders — unmatched in large-scale identification.
Official NIST FATE results score facehawk-001 at a GINI of 0.065 for age estimation — exceptional demographic consistency — with a MAX–MIN variability of 2.38 years across regions.
0.065
GINI — age estimation
2.38 yrs
Max–min variability, all regions
GINI across NIST benchmarks — lower is fairer:
Age estimation shows even tighter demographic consistency than 1:N identification — near-perfect parity across every region tested.
Scale 0 – 0.9 · industry average 0.6 – 0.8
Scale 0 – 11 KB · industry average 5 – 10 KB
Scale 0 – 220ms · industry average 100 – 200ms
Scale 0 – 550 MB · industry average 300 – 500 MB
Fairness is not an afterthought. It is the foundation.
169 MB
Config data
35 MB
Library data
343 MB
Template memory
2,070 B
Template size
362 ms
Generation (mugshot)
1,465 ns
Comparison (mate)
1,475 ns
Comparison (non-mate)
The compact template size and efficient comparison times make facehawk-001 suitable for embedded and edge deployments where bandwidth and storage are constrained.
facehawk-001 participated in NIST's Twins disambiguation testing, demonstrating capability in one of facial recognition's most challenging edge cases.
| Metric | Result |
|---|---|
| Mugshot FNMR | 0.0440 |
| Template failure rate | 0.2720 |
| Identical twins FMR | 0.541 |
| Fraternal same-sex FMR | 0.384 |
| Fraternal diff-sex FMR | 0.0000 |
While identical twins remain a challenge for all algorithms, facehawk-001 shows strong discrimination for fraternal twins and non-twin comparisons.
Measures how often a comparison of images from two different people produces a similarity score at or above threshold — the "false accept" rate.
Measures how often a comparison of two images of the same person produces a score below threshold — the "false reject" rate.
NIST evaluates across multiple datasets to simulate real-world operational conditions:
Controlled, frontal, cooperative capture.
Document photos with varying quality.
Uncontrolled capture environments.
Self-service capture with variable pose and lighting.
The "ΔT ≥ 12 Years" scenario is particularly demanding, testing algorithm robustness to aging.
Full validation detail, including the complete NIST report, shared under NDA.