Microsoft Kinect Heats Race Debate: Does Face-Recognition Software Discriminate?
The U.S. has made many strides over the past century toward treating people with equality regardless of the color of their skin, but it seems that technology, surprisingly, has moved at a slower pace. The launch of Microsoft's (MSFT) long-anticipated motion-sensing video game, Kinect for Xbox 360, has fueled the debate over whether facial-recognition software discriminates against some races.
GameSpot noted in a review Wednesday that Kinect's facial recognition capabilities appear to fail to recognize players with dark skin. It's by no means the first time that facial-recognition software has faced allegations of being less than fair and equal fashion. Remember last year's brouhaha over claims that HP's (HPQ) Webcams failed to track people with dark skin?
"It's unfortunate to have these incidents in the creation of a product," says Patrick Flynn, a professor of computer science and engineering at the University of Nortre Dame and a biometrics research who has authored papers on face recognition technology. "There's nothing like the harsh light of publicity to point out a need for a remedy."
Here's what GameSpot had to say in its review of the Kinect:
While testing out the Kinect, two dark-skinned GameSpot employees experienced problems with the system's facial-recognition abilities. The system recognized one employee inconsistently, while it was never able to properly identify the other despite repeated calibration attempts. However, Kinect had no problems identifying a third dark-skinned GameSpot employee, recognizing his face after a single calibration. Lighter-skinned employees were also consistently picked up on the first try.
Diversity Training for Software
While Flynn, who specializes in biometrics for government applications, noted that he's unfamiliar with the development process behind Microsoft's Kinect and other commercial products using face-recognition software, he said the difficulty with face-recognition software in general is that it usually needs to be trained to adequately detect various people.
Similar to voice-recognition software, in which accuracy improves the more it's exposed to a person's voice, face-recognition software works better with increased exposure to a person's face. "As a rule, it's been difficult to have a sufficient number of dark-skin people to provide the data needed to train these systems," Flynn says.
For one thing, it's expensive to recruit subjects, so companies often turn to "locals," such as their own employees and families, instead, he says. But "the demographic balance in the high-tech industry is not very representative of the country as a whole," Flynn says, pointing to a San Jose Mercury Newsarticle on diversity at Silicon Valley tech companies.
Another way to obtain facial data is to use databases assembled by other organizations, he says. "Some of these database are also lacking in demographic balance," he adds. "In some cases this is unavoidable on a practical level."
Why is it so expensive to build a diverse database pool? Researchers ideally need people to commit to having their photos taken numerous times over the course of days, weeks and preferably years, says Damon Woodard, an assistant professor at Clemson University and director of its biometrics and pattern recognition lab. And folks who are willing to be part of a data pool typically expect some form of compensation, he adds.
This data is critical when the face-recognition software is in the development stage. The greater the diversity of the people in the data pool, the better the chances it will work with the market it's designed to address. And diversity in the training process is especially important because most of these systems can't generalize. For example, it's proven extremely difficult to develop systems with flexible enough learning capabilities that they recognize eye shapes of people from Europe and not just of people from Northern Europe, Flynn says.
Further Investigation
Microsoft, for its part, is investigating the problems cited in the GameSpot report, according to an email statement. "We are investigating GameSpot's problems with the facial recognition feature of Kinect for Xbox 360. Kinect will work with people of all shapes and ethnicities," the company says.
When the HP Webcam issue arose late last year, Consumer Reports issued its own report finding that the problem had to do with inadequate lighting: One solution was to shine more light onto the face of users with dark skin.
And in the case of Microsoft's Kinect, Consumer Reports also found no problems in its tests. In a blog post late Thursday, it said:
The Kinect recognized both players at adequate light levels and failed to recognize both players at levels that were too low. At no time did it recognize one player and not the other...
....To sum up, the Kinect recognition issues affect automatic logins to the Xbox it's attached to if it's being used in poorly lit environments. But even without facial recognition, players can still log in to their personal avatars using simple hand-gestures or with standard Xbox controllers.
Regardless of whether the test from Consumer Reports or GameSpot proves to be more accurate, experts Woodard and Flynn both say they'd like to see more diversity in the data pool used to develop face-recognition software.
Get info on stocks mentioned in this article:
