The Face Recognition Vendor Test 2002 (FRVT 2002) looked at 10 companies' work on face recognition and said they had made "significant advances" on the state of the art.
"The performance results...show an improvement in the capabilities of the face-recognition systems over the last two years," concludes the report released this month. "On comparable experiments in...2000, there has been a 50 percent reduction in error rates."
The team that produced the report included a slew of government agencies. Evaluators included scientists from the U.S. Department of Commerce's National Institute of Standards and Technology (NIST), which issued the report; the Defense Advanced Research Projects Agency (DARPA) and the Defense Department's Counterdrug Technology Development Program Office. Other sponsors of the report include the Justice Department's National Institute of Justice, the State Department, the Transportation Security Administration and the FBI.
NIST said it had performed the study in response to two federal laws--the USA Patriot Act and the Enhanced Border Security Act--that were passed after the Sept. 11, 2001, terror attacks.
The study, which matched 121,589 images of 37,437 people drawn from the State Department's Mexican nonimmigrant visa archive, evaluated how well the various commercially available systems verified identity, identified unknown faces and detected people on a "watch list."
The systems were able to verify a person's identity 90 percent of the time, with a 1 percent error rate. That's about as well as 1998 vintage fingerprinting technologies, NIST said.
The new study looked at demographic factors for the first time. Men were easier to identify than women, and older people easier than younger.
Another crucial variable turned out to be whether people were being identified indoors or outdoors. Indoor environments, where lighting is more easily controlled, provide results that are twice as reliable.
In a mandated report to Congress on the test results, NIST recommended that face-recognition technologies be combined with fingerprinting technologies for greater accuracy.
The vendors evaluated in the study were AcSys Biometrics, Cognitec Systems, C-VIS Computer Vision und Automation, Dream Mirh, Eyematic Interfaces, Iconquest, Identix, Imagis Technologies, Viisage Technology and VisionSphere.
The report singled out a technology called three-dimensional morphable modeling, which fits a 3-D model to a facial image taken from the side, and then turns that image into a frontal shot. NIST said this technology could "significantly improve nonfrontal face recognition."
The accuracy of face recognition from video sequences, by contrast, was only marginally better than from still images, said NIST.
NIST said it had developed an XML-based evaluation protocol and scoring suite for the report that would be useful for both face-recognition tests and other evaluations of biometric technologies.