NIST conducting program to assess facial recognition accuracy
The U.S. National Institute of Standards and Technology announced a new program called the Face in Video Evaluation (FIVE), which is designed to assess the capability of face recognition algorithms to accurately identify or ignore people captured on video.
The tests will assess both comparative and absolute accuracy measures, working toward the end goals of figuring out which algorithms are most effective.
Additionally, the tests will assess whether any of these algorithms can be applied to the primary operational use-cases of a high volume screening of people in the highly-populated spaces, low volume forensic examination of footage from a crime scene, for people in business meetings, and for people appearing in television footage.
NIST said these applications differ in their tolerance of false positives, whether a human examiner will review outputs, the prior probabilities of mate compared to non-mate presence, and the cost of recognition errors.
The FIVE program is the latest in a series of face recognition vendor tests performed in 2000, 2002, 2006, 2010, and 2013, all of which provided quantitative statements of accuracy and speed of predominantly still-image face recognition algorithms.
The API publication and public comment period will commence August 15 and run through October 30, followed by phase 1 of the algorithm submission period which runs from November 17 to January 8, phase 2 of the submission period which occurs in spring 2015, and phase 3 which runs throughout summer 2015.