My print column this week examined facial-recognition software, and the confusing numbers about its accuracy. A recent study seemed to suggest that the technology has become highly reliable even at identifying people from a pool of Facebook photos, but a close look at the results shows they are less impressive than they seem.
The U.K. police have learned as much when they tried using such software to identify people filmed by surveillance cameras during the riots last month. A person familiar with the matter said the police only see it as one weapon in the arsenal and not a major one, at that. That is because the success is dependent on already having a picture of a suspect on file, and for the CCTV image to be a good one. Half a face, or a blurry one -- which many of these images are -- won't do, this person said. Relying on human identification, from the public and the police, remains the best way to identify suspects.
Simon Lubin, a spokesman for the British Transport Police, said the department has tried the software in the past but declined to confirm whether it was being used in the riots. The trouble, he said, is it isn't yet very good at picking out one person in a crowd of thousands. "You've got thousands of people going through the top of an escalator" in London's underground train system, Mr. Lubin said. "It's not sophisticated enough yet to be operationally worthwhile. That may change. Technology is changing all the time."
For now, experts agree, the technology has a lot of trouble with imperfect images. "Facial recognition only ‘works' (depending on what your criteria are) under optimal acquisition conditions," Arnout Ruifrok of the Netherlands Forensic Institute wrote in an email. "As soon as pose, lighting and occlusion are included, the performance quickly deteriorates."