NIST launches new face recognition evaluation
31 October 2017 06:16 GMT

The National Institute of Standards and Technology (NIST) has launched a major new evaluation of face recognition technologies.

FRVT 2018 will measure advancements in the accuracy and speed of one-to-many face recognition identification algorithms searching enrolled galleries containing at least 107 identities. 

Starting in February 2018, NIST told Planet Biometrics that FRVT activity aims to:

  • Estimate absolute accuracy on databases with N exceeding 20 million identities.
  • Give a robust comparison of leading commercial and academic algorithms, in terms of accuracy and speed.
  • Give performance estimates appropriate to operator-led investigative searches where an analyst will launch a search and then adjudicate hypothesized identities from an automated algorithm.
  • Give performance estimates appropriate to higher search volume applications where a matching threshold must be instituted to suppress false positives such that the automated system produces candidates at a rate suited to the available labor supply.  This situation applies in duplicate checks for passport, visa or driving license applications.
  • Measure accuracy as a function of demographic effects, and image quality degradations. 
  • Measure accuracy as a function of time lapse (ageing).
  • Evaluate using non-cooperative, less constrained, images also.  This will follow the 2017 IARPA sponsored Face Recognition Prize Challenge the report of which NIST expects to release in November.

"Given the revolution in FR technologies over the last five years, we seek to capture potentially large advances in performance for, primarily, the application of the technology to cooperative portrait images,"  Patrick Grother, a computer scientist specializing in biometrics at the National Institute for Standards and Technology, told Planet Biometrics.

The first draft of the C++ API for the FRVT 1:N 2018 evaluation is now online.  The document, which also describes the goals and concept, is open for public review.  

Please submit suggestions, comments and questions to by November 15, 2017.