NIST face biometrics accuracy update shows same top vendors for people wearing masks

The National Institute of Standards and Technology (NIST) has updated the results of its testing for the accuracy of facial recognition algorithms identifying people wearing masks with a group of entries submitted since the agency issued a report on the topic in July.

The overall results look roughly the same, as the top two results, from Deep Glint and Paravision, are from the same biometric algorithms, which were each originally submitted in 2019.

NIST has also updated its latest report for FRVT 1:1 Verification, the first to make it mandatory for all algorithms submitted to any track of FRVT to be compiled for the CentOS 8.2 operating system.

The results have also been updated to list the most accurate facial recognition developers instead of algorithms, choosing the most accurate one from each developer based on visa and mugshot results.

Moreover, five new algorithms were added to the FRVT 1:1 Verification report since September 18, 2020: Aigen, Cortica, Kookmin University, Securif AI, and VinAI.

The new report also includes results from Fujitsu Laboratories, Hengrui AI, and X-Forward AI, all developers who had previously submitted algorithms.

A few companies stand out for high accuracy results in the reports.

Deep Glint also ranked first in the FRVT 1:1 test. The company has been growing steadily in the past year, and its facial biometrics have been used by Chinese police to catch criminals.

VisionLabs, on the other hand, ranked second in the FRVT 1:1 Verification report, and fourth in the FRVT Face Mask Effects one. The company has recently updated its algorithms to deploy contactless biometric payment solutions in Russia.

Ranking third in both reports was a new algorithm by developer xforwardai-001, while Vocord finished seventh in the FRVT 1:1 Verification report, and ninth in the FRVT Face Mask Effects one.

NIST said that after these new reports, the Institute intends to continue evaluating algorithms on various mask datasets. Future testing will include algorithms developed since the pandemic began specifically to work with masks. For this test, like the first, NIST used digitally-generated masks, rather than real ones.

To allow algorithms to evaluate both masked and unmasked faces, NIST has increased the amount of time allowed to extract facial features from 1.0 to 1.5 seconds.

Article Topics

accuracy  |  algorithms  |  biometric identification  |  biometrics  |  Deep Glint  |  facial recognition  |  NIST  |  Paravision  |  testing  |  VinAI  |  VisionLabs  |  Vocord