Facebook’s AI research group has just developed a new piece of software that can tell whether photographs of human faces are the same. The so-called DeepFace algorithm matches faces like humans. The paper with the telling title ‘DeepFace: Closing the Gap to Human-Level Performance in Face Verification’ will be presented at the Conference on Computer Vision and Pattern Recognition (CVPR) later this year.
As MIT Technology Review’s Tom Simonite correctly points out, one must distinguish between facial recognition and facial verification:
DeepFace performs what researchers call facial verification (it recognizes that two images show the same face), not facial recognition (putting a name to a face),
Tom Simonite, MIT Technology Review
DeepFace is a portmanteau, made of the terms ‘deep learning’ and ‘face’ (verification, if you will). Deep learning is used for recognizing patterns in large data sets. Deep learning belongs to machine learning and has been applied to speech recognition (for example, Siri) and computer vision.
DeepFace uses deep neural networks. Computer models that mimic a nervous system and its connections, that is. DeepFace’s network has over 120 million connections of these virtual nerves. The network must be fed data, lots of it. That, of course, is Facebook’s least concern, as it owns the world’s largest photo database. For the tests, though, it used a subset of four million photographs.
What Makes DeepFace Different?
DeepFace works different than current facial recognition software. Existing facial recognition software usually analyzes flat images of people’s faces, measures and quantifies physical features such as the distance between the eyes. These algorithms usually fail when the image is flawy, for example, when it contains harsh light and shadows, or when the face has not been directly looking into the camera. By the way, this is the reason why need to follow all these guidelines when shooting photographs for our passports – because otherwise they can’t be recognized.
Now, DeepFace takes flat photographs, glues them onto a 3D model and ‘frontalizes’ them. In other words: the 3D model is turned until the face looks straight to the camera. This process is called normalization. It ensures that all photographs have the same starting point before being analyzed and matched against each other.
Then, the deep learning analysis takes place: the face is being numerically quantified and described and eventually compared against the other images’ numbers in the database. By the way: DeepFace correctly matches faces 97.25 percent of the time. Humans score only minimally better: 97.53 percent of the time. That’s pretty good, isn’t it?
What Will DeepFace Be Used For?
That’s indeed a very good question. Facebook (the company) has announced it will not integrate DeepFace in Facebook (the software) anytime soon and does not mention other applications in the paper. Better still, I’d say. This leaves room for imagination. I think DeepFace might be used for:
– Optimizing CCTV and tracking terrorists
– Simplifying biometric authentication
– Loosening passport photograph guidelines
What’s your take on it? Any ideas? Let me know via Twitter! https://www.twitter.com/martinangler