Wilson Dizard reports on NIST’s testing of 14 facial recognition systems.
The test had three parts. First, NIST asked the systems to match a facial image against the database of images and find 10, 20 or 25 similar images. Next, the systems had to verify identities using the database of images. Finally, NIST checked each system‚s reliability under different lighting conditions and monitored the speed of each application.
The three top-rated systems verified identities correctly 87 percent to 90 percent of the time with a false-alarm rate of 1 percent. When NIST specified a false-alarm rate of 0.1 percent, the success rate dropped to between 79 percent and 82 percent.
When checking facial images against a watch list of 25 images at a false-alarm rate of 1 percent, the top three systems were accurate about 80 percent of the time. The success rate fell to below 60 percent when NIST expanded the watch list to 3,000 images at the same false-alarm rate.