Paravision takes top spot in accuracy matching to NIST’s new high-angle dataset
An algorithm from Paravision has been found most accurate in the newest category added to the National Institute of Standards and Technology’s Face Recognition Vendor Test 1:1 Verification.
The NIST FRVT 1:1 added the ‘Visaborder Yaw≥45 degrees’ category was added for the agency’s February 9, 2023 update. The dataset is described as a “new set of non-frontal portrait to border comparisons.” Paravision’s latest algorithm was found in the June 16 update to have a false non-match rate (FNMR) of 0.0025 percent with the false match rate (FMR) set to 0.000001 percent. A pair of algorithms from Chinese developer Cloudwalk were next-most accurate, followed by Paravision’s previous submission.
The company also notes in its announcement a steep drop-off in the accuracy of algorithms tested against the new database after the top five.
Paravision also claims to have scored the top overall result among biometrics developers from the U.S., UK and EU. The company’s algorithms were found among the five most accurate in two mugshot datasets and the kiosk category, and within the top 10 on two more datasets related to borders.
“In evaluation after evaluation, we’ve seen Paravision technology and products ‘Built with Paravision’ judged as the best performing globally. This particular test is a good indicator as to why: We do extremely well on high quality images while also delivering outstanding performance when conditions get tough, users aren’t experts, or user experiences are less constrained,” says Charlie Rice, CTO of Paravision.
NIST also breaks down the results by different demographics, with the latest evaluation showing Paravision’s worst case single-demographic differential is substantially lower than all competitors, according to the announcement.
The latest algorithm submitted by Paravision will be a part of the company’s forthcoming Gen 6 facial recognition software, according to the announcement. Paravision’s Gen 6 technology is scheduled for delivery to customers this summer.
The developers found to be most accurate in the evaluation remain broadly similar to recent updates, but with the notable addition of Recognito and Viante.AI. Both are based in the United Arab Emirates, and they joined the evaluations with submissions on May 24 and June 9, respectively.