Facebook's Facial Recognition 'Approaching Human-Level Performance'

Facebook has been working on facial recognition for years to auto-tag photographs, but has now reached a point where its technology is 'closely approaching human-level performance.' In fact, in some ways it might even be better.

The social network's in-house research team has been working on a project called DeepFace: cutting-edge facial recognition software which maps 3D facial features, turns them into a flat model, then filters the result by color to identify specific facial elements.

And boy, does it work. Pitted against a test where it's asked to identify whether two people in side-by-side images are the same person, it's 97.25 percent accurate. When humans takes the test, they score 97.5 percent. Good grief that's impressive. It's also taken some time and patience to achieve. The system uses a staggering 120 million parameters to identify faces, and has been trained using a pool of 4.4 million labeled faces from 4,030 different people on the social network to reach this point.

The feature won't be rolled out on the website just yet, though. In fact, Facebook is taking it to the IEEE Conference on Computer Vision and Pattern Recognition this June to see what academic and other researchers make of it. Hopefully more than a few privacy advocates as well. [Facebook via Technology Review]