Face Recognition Software that can Scan Through Millions of Faces Per Second

I received a couple of emails today about a face recognition system from Hitachi Kokusai Electric that can search up to 36 million faces per second. One of the impressive parts is that it can recognize a face with 30 degree deviation from the camera and the small image size required 40 x 40 pixels. Like almost all commercial face recognition systems, this requires a man in the loop to verify the images. The article doesn't give the accuracy, but usually for a face recognition system to be useful there needs to be an accuracy high enough for a person verifying the images to not be overwhelmed with images of suspects.

Errors in face recognition are generally caused by three factors: pose errors i.e. the person having a there head in a different position than the stored image or training image of the person, lighting issues i.e. the lighting on the person is different than the training image, and a catch all called non-rigid transforms. Non-rigid transforms includes most changes in a person's appearance from the stored image of the person to image one is trying to match to the image. These include changes in skin color like a person getting a tan from a day at the beach or have pale skin because they are sick, as well as, occlusions to face like sunglasses, scarfs, and the new phenomenon of face recognition dazzle. I talk about these errors more in an article I wrote about face authentication and in my PhD dissertation.

I'm a little dubious that this software will be as effective as it claims. While fixing pose and lighting errors are almost solved problems by using 3D modeling of heads to artificially replicate the pose and lighting, the non-rigid transforms still pose a problem. So, this system may not work if you get a tan or put on sunglasses.