Identical twins -- the ultimate test for biometrics

Published 29 March 2011

Researchers recently concluded that facial biometric technologies were not accurate enough to distinguish between pairs of identical twins; researchers took photos of over 126 pairs of identical twins in a variety of conditions of varying quality to provide different test conditions for biometric scanners; the photos were tested against three of the highest performing facial recognition and found that under real world circumstances the systems could not accurately distinguish twins; researchers recommend calibrating facial recognition algorithms to analyze minute facial characteristics as well as high-resolution photos to increase accuracy

Researchers recently concluded that facial biometric technologies were not accurate enough to distinguish between pairs of identical twins.

At the annual Twins Days Festival in Twinsburg, Ohio, University of Notre Dame professors Kevin W. Bowyer and Patrick J. Flynn sought to document as many pairs of identical twins as they could find.

The Twins Days festival is particularly noteworthy for the fact that twins go at great lengths to look as indistinguishable as possible, going so far as to match haircuts, outfits, and jewelry.

Bowyer and Flynn took photos of over 126 pairs of identical twins in a variety of conditions to provide different test conditions for biometric scanners.

The researchers then scanned the photos on three of the highest performing facial recognition programs as determined by the National Institute of Standards Technology, a federal agency that tests biometric systems against mug shots, visa photos, and other government databases to determine their accuracy.

The three biometric systems performed well in ideal circumstances and were able to distinguish twins fairly well when analyzing high resolution photos under studio lights with neutral expressions.

Using what they called an “equal error rate metric” based on how likely a program was to make a false positive match as it was to make a false negative mismatch, the researchers found that the three systems all had an equal error rate of less than 5 percent.

But, when the researchers tested the technology under real-world conditions, the programs performed poorly.The systems were prone to producing errors when comparing pictures of smiling and neutral faces or photos taken under ambient light rather than studio lights.

To further simulate realistic circumstances, Bowyer and Flynn returned to the Twins Days festival a year later and tracked down twenty-four of the twins that they had originally photographed. The researchers took photos once more, but this time in ambient light, and tested the software’s ability to distinguish between the photos taken a year apart.

It’s as if you have a nice indoor studio image of a man, and then you just happen to catch a shot of him a year later as he’s walking down the street. In that circumstance, the best-performing algorithm had about a 17 or 18 percent equal error rate. That’s awful for almost any application,” Bowyer summarized.

Drawing lessons from the study, Bowyer recommends that facial recognition software analyze minute characteristics rather than broader features like the distances between facial features, which most systems use.

Even identical twins are distinguishable, but it’s a collection of minute differences: a mole here, a crooked tooth there. You can only tell the difference between the two faces if you have reliable detection of fine details,” he said.

Bowyer also suggests using higher resolution photos that actually show these smaller distinguishing features.

The two researchers presented their findings on 22 March 2011 at the IEEE International Conference on Automatic Face and Gesture Recognition.

Their research was sponsored by the FBI which is actively seeking to improve biometric facial recognition technology, which is currently inadmissible as evidence in court due to its inaccuracy.

view counter
view counter