Facebook has gotten so good at recognizing people from images that it doesn’t even need to see their faces anymore.
At the Computer Vision and Pattern Recognition conference in Boston earlier this month, the social network presented research that shows it can identify individuals with 83% accuracy, according to The New Scientist.
The key is relying on context. “There are a lot of cues we use. People have characteristic aspects, even if you look at them from the back,” Yann LeCun, Facebook’s head of artificial intelligence, said at the conference. “For example, you can recognize Mark Zuckerberg very easily, because he always wears a gray T-shirt.”
Facebook has been using its facial recognition software to suggest tags for uploaded photos since 2010. But it isn’t the only one that’s been using computer vision to make photo tagging more seamless.
In May, Flickr debuted new auto-tagging capabilities as part of its redesign. Though the technology helped pull out otherwise unseen and untagged images, it wasn’t perfect. “Sometimes a bicycle looks like a motorcycle,” Andrew Stadlen, Flickr’s director of product management, told Quartz shortly after the redesign. “Sometimes grandma looks like a cat.”
Humorous as that might be, there are times when getting it wrong can be much more offensive. Like when Flickr started mistagging black people as apes and concentration camps as jungle gyms. Indeed, this is one of the big challenges of taking something that originated in an academic setting—in this case, computer vision to predict the subject matter of photos—and turning it into a consumer-facing product.
Facebook hasn’t said when or how it’ll deploy the new technology to recognize people when their faces have been obscured. (It did recently release a new standalone photo app called Moments.) But there’s one lesson it can take away from Flickr: It should take pains to train users to understand the limits of its technology. “People have extremely visual systems. Computers are not yet as good as humans,” said Flickr research engineer Pierre Garrigues. “To make this possible, we want users to be comfortable with some level of error.”