[bsenftner]

Yes, it exists, and I use it every day. In fact, I have a partnership with the DOD/NSA subcontractors who created it, and I have applied the technology to the automated creation of 3D Avatars given a single photo of a person's face. My version of the tech is available at www.3D-Avatar-Store.com. The system works as follows: a person's head is laser scanned, at the same time dozens of single photos are taken of that person from different angles, different lighting conditions, and different quality cameras. Then a neural net is trained to associate each photo with the laser scan data. After a few thousand trainings, anyone's single photo would generate a reasonable likeness of them in 3D. Today, after over a decade of training and 10's of thousands of scans added to the training database, we get remarkably good quality 3D reconstructions given a single photo of anyone, any age, any ethnicity. There are limitations, such as the face needs to be visible, and it needs to be within +/-30 degrees of facing the camera for the 3D reconstruction to be suitable for facial recognition. And that is the point of the original application of this technology: a facial recognition pre-processor. It corrects the face angle and removes any facial expression, creating a likeness of the face in a passport style photo perfect for facial recognition. But, of course, I'm not using it for that. I use it to auto-magically create 3D avatars for video game studios.

Oh, it's really, really fast. My current system is based upon two generations back of neural nets, and they reconstruct from one photo in 0.9 seconds. The latest generation does 144 reconstructions simultaneously given HD video feeds.

[wicknicks]

Very interesting technology. I used your website, uploaded three pictures -- it failed on the first, and worked somewhat on the others. I'm guessing it needs a more experienced user base? Assuming that model generation requires human supervision, do you think the neural network algorithm scales to the numbers mentioned in the article (trained on 12 million photos, and operating in 7 different states)?

BTW, all the best with your website. Hope you take more avatars "where go where no man has gone before!"

[anonymousDan]

What's the state of the art in terms of extending such systems to handle faces that are partially obscured (e.g. wearing glasses)? References to further resources in the area accepted and appreciated! I'm going to start work on a face recognition system soon and I'm keen to learn more about their current limitations.

[bsenftner]

It's just an issue of training. I'd expect a similar setup as I describe could be used to associate a series of faces with and without different things attached to their face, and a neural net could be trained to generate, for example, a no eye glasses wearing image from one with eye glasses. We already do that with facial expressions. I think the issues to be improved upon are extending the recognition beyond the face - include ear characteristics to extend recognition beyond +/- 30 degrees, and although my lab tells me analyzing gait is a dead end, I feel there might be something there that can help. Also lighting neutralization is an area that needs some help, which includes illumination balancing without removing important details, shadow removal, and colored illumination white balancing.