NIST evaluates face recognition identification algorithms
23 November 2018 16:02 GMT

The National Institute of Standards and Technology (NIST) has announce the release of NIST Interagency Report 8238, which documents its first large scale evaluation of face recognition identification algorithms since a similar trial in 2014.  

NIST has related that the primary results are that the new generation of face recognition algorithms – those based on convolutional neural network architectures – give massively reduced error rates compared to the prior generation.

The most accurate algorithms now give a factor of twenty fewer errors than the best algorithm we reported in 2014.  This constitutes an industrial revolution. Whether the improvements now move into a more evolutionary phase may become apparent in a further round of tests recently underway at NIST.

The industry has expanded at least threefold since 2014, when just 16 developers participated in NIST’s FRVT evaluation. In 2018, more than 50 developers have submitted algorithms to NIST, although only 45 completed the tests.

There remains a broad spectrum of capability across the industry – while algorithms from 28 developers are superior to the best algorithm documented in the 2014 report, the leading algorithms offer much lower error rates.

The NIST results are likely to drive a round of upgrades by end users, as developers make the new generation prototypes evaluated at NIST available in their product offerings.