Facebook almost as good as humans at recognising faces

When you are called Facebook, this is a no-brainer. The social network’sresearchers have built DeepFace, an algorithm that can pick a face out of a crowd with 97.25 per cent accuracy. That means it is almost as good as we are at recognising a face.

The trouble with most face recognition systems is that if the faces are slightly off-centre in the photo the software will struggle to find a match. Yaniv Taigman and colleagues at Facebook’s AI lab found a way round this problem. They created a 3D model of a face from a photo that can be rotated into the best position for the algorithm to start matching. They then used a neural network that had been trained on a massive database of faces to try and match the face with one in a test dataset.

The team achieved 97.25 per cent accuracy on the test dataset that contained more than 13,000 images of faces collected from the web.

That falls just fractionally short of the best that humans can do: a test using real people recruited on Amazon’s Mechanical Turk crowdsourcing site managed 97.5 per cent.

Facebook’s neural network was trained on a dataset of more than 4 million images, containing more than 4000 separate identities, each one labelled by humans. The researchers don’t say where this massive dataset of faces came from, but it’s safe to assume that Facebook users are the source.

Details will be presented at the Computer Vision and Pattern Recognition conference in Columbus, Ohio, in June.

Gary Huang of the University of Massachusetts in Amherst says that a Chinese machine vision start-up called Face++ claims to have a comparable level of face recognition accuracy. Face++ is used on China’s largest online dating website, Jiayuan.com, where you can upload a photo of someone you find attractive, and the website will find users that look similar. Huang says DeepFace or Face++ aren’t accurate enough for a courtroom, but could be very useful for automatically finding matches to surveillances images, with all the privacy concerns that come with that.

What could Facebook do with this level of face recognition? “They could identify every one of your friends in all of your photos,” says Christopher Kanan of the California Institute of Technology. “This would let them know who your real friends are – people you actually hang out with in the real world.”