Population-scale face recognition
Identifying individuals by face against galleries of tens of millions of people from cameras in public and private spaces.
State of the art
Top NIST FRTE 1:N algorithms reach ~0.05% FNIR on mugshots and ~1% on harder imagery at 12M gallery size (2026).
- Accuracy drops for low-quality, off-angle, masked or aged images
- Demographic differentials remain for some algorithms
Who’s working on it
- Patrick GrotherResearcherNational Institute of Standards and Technology
Runs the FRTE/FATE evaluations; his 2025 analyses show 1:N mugshot accuracy "hitting a wall" while legacy algorithms still miss 28% at FMR 1e-6
Continuous open benchmark on sequestered galleries up to millions of identities; the de facto global accuracy standard, with dozens of vendors submitting every four months
- NECCompany
Supplies border-control and national police identification systems worldwide and ranks at the top of NIST 1:N leaderboards (NEC-2023-12 onward)
- IDEMIACompany
Top-ranked 1:N vendor supplying civil identity and border systems; its 2024 submissions anchor NIST's cross-age accuracy results
- Clearview AIStartup
Built a gallery of billions of scraped web images and submits competitive algorithms to NIST (Clearview AI-2024-02), making it the main non-state scale case
- HikvisionCompany
Sells face recognition as a standard feature of commodity cameras in nearly every country, which is what makes the capability ubiquitous rather than elite
- Paravision ↗Company
Highest-ranked US/European 1:N vendor in NIST's January 2025 evaluation, with edge and mobile SDKs for deployment outside data centers
What to invest in
| Kind | Investment | Scale |
|---|---|---|
| research | Independent bias and accuracy audits for in-the-wild video | $ < $10M |
Depends on
A root node: no prerequisites modeled.