NIST FRVT (Face Recognition Vendor Test)
NIST FRVT is the Face Recognition Vendor Test, an ongoing evaluation run by the US National Institute of Standards and Technology that measures submitted face recognition algorithms on sequestered datasets. NIST reports results per algorithm and dataset; it does not endorse vendors and issues no certification.
What does NIST FRVT actually measure?
Participants submit an algorithm as a compiled library. NIST runs it on datasets the participant has never seen, under fixed conditions and on NIST hardware, and reports error rates by dataset, by task and by demographic group. Nothing is self-reported, which is what makes the results comparable between vendors.
- 1:1 verification: error rates on visa, mugshot, border-crossing and wild image sets.
- 1:N identification: search accuracy against galleries ranging up to millions of enrolled identities.
- Demographic effects: error rates broken down by sex, age band and country of birth.
- Presentation attack detection and morph detection, evaluated in separate tracks.
- Template generation and search time, measured on the same hardware for every participant.
Why can no company be NIST certified?
NIST is a measurement institute, not a certification body. Its evaluations produce reports describing how an algorithm performed on specific data at a specific time. There is no pass mark, no certificate and no approved-vendor list, so a claim of being NIST certified or NIST approved describes something that does not exist.
What a vendor can legitimately say is which algorithm they submitted, when, and where it placed on a named dataset. Anything phrased as an endorsement is a claim NIST does not make about anyone.
How should a buyer read an FRVT report?
- Find the exact algorithm identifier, not the company name: vendors submit many algorithms and results differ between them.
- Check the submission date, because a leading result from several years ago has been overtaken repeatedly.
- Read the dataset that matches your conditions: visa-quality portraits predict nothing about a wide-angle camera in a corridor.
- Read the demographic breakdown rather than the aggregate, because the aggregate hides the variation that matters operationally.
- Compare rank on your dataset, not overall, since no algorithm leads every track.
A rank is evidence about an algorithm under laboratory conditions, and it remains the best independent evidence available in this field. It is not a prediction of what a system will do on a specific site with specific cameras, which is what a pilot measures.
Frequently asked questions
- What is NIST FRVT?
- NIST FRVT is the Face Recognition Vendor Test, an ongoing evaluation by the US National Institute of Standards and Technology that measures submitted face recognition algorithms on sequestered datasets under fixed conditions. Results are published per algorithm, dataset and demographic group, and NIST does not endorse vendors.
- Can a company be NIST certified for face recognition?
- No. NIST is a measurement institute, not a certification body: it publishes evaluation reports rather than certificates, and maintains no approved-vendor list. NIST certified and NIST approved describe something that does not exist, so treat either phrase in marketing material as a reason to ask for the actual report.
- Is participation in FRVT voluntary?
- Yes. Vendors choose to submit algorithms, and the results are published whatever they show. That is what gives the programme its value as independent evidence: a participant cannot submit, see the outcome and then withdraw it from publication.
- Does a high FRVT rank mean a system will work at my site?
- It means the algorithm performed well on that dataset under laboratory conditions. Site performance depends on camera placement, resolution across the face, lighting, throughput and gallery size, none of which the evaluation covers. A rank narrows the shortlist; a pilot on your own cameras produces the number that describes your deployment.
- How often is FRVT updated?
- The programme runs continuously and publishes updated reports as new submissions are evaluated, rather than on a fixed annual cycle. This is why a result should always be read with its date attached: a leading position from an earlier report may have been overtaken many times since.
- Does FRVT measure demographic differences?
- Yes. NIST reports error rates broken down by attributes including sex, age band and country of birth, and has documented that the size of demographic variation differs substantially between algorithms. Those breakdowns, rather than the aggregate figure, are the part worth reading in a procurement review.