Tech from University of Nottingham can create a 3D face model from a single photo

All photos courtesy the author.

I remember walking with a friend through downtown San Mateo, California (part of the Silicon Valley) years ago, and seeing a store that offered to scan your body with a giant 3D scanner. They would then produce a full-color, action figure sized version of you for around $50.

My friend suggested a delightful prank: print yourself en-masse, and then send the tiny action figures to all your friends and family. The figures would be high enough quality that people would probably feel compelled to keep them for a while. And thus there you’d be — in creepy, miniature form — haunting the desks and junk drawers of all your loved ones until they mustered the chutzpah to throw you away.

At the time, that would have been prohibitively expensive. But the combination of a new 3D face mapping technology and affordable 3D printing is bringing that hilarious and slightly horrifying vision closer to reality. It’s also creating tech which has important applications in video gaming, virtual reality, and even security.

The tech, called Large Pose 3D Face Reconstruction from a Single Image via Direct Volumetric CNN Regression, was developed by researchers at the University of Nottingham and Kingston University, both in the UK. It allows users to upload a single photo of their face, from which the tech is able to derive a 3D facial model.

Make no mistake, this is a deviously hard tech challenge. Usually, creating 3D models from photographs–especially for something as complex as a face–requires lot of photos, taken from multiple angles, with excellent lighting.

As report in the Verge, however, the research team behind the tech managed to cut all this out by training a deep learning model using high-quality 3D facial models and their corresponding 2D images. The resulting neural network can take in a 2D photograph and quickly derive a passable 3D model of the subject, no giant walk-in scanner required.

It’s a technology that’s fast being applied by others. Ubiquitous Chinese facial recognition company Face++ performs a similar function, but they charge about $2 per facial model generated.

The best thing with both processes, though, is that the end result is an STL file, which is an industry standard 3D model file which can be used with most 3D printers. So if you use the tech to make a model of your own face and you have a Makerbot (or an account with a company like Shapeways), you can quickly create a 3D bust of your own face for minimal cost.

I decided to test this out, and uploaded a photo to the University of Nottingham’s platform. In a few minutes, as promised, it returned a 3D model of my face as an STL.

My face, rendered by the University’s tech.

I then imported this into Makerbot Print, fiddled with the settings a bit to get it flat-ish on the build plate, and set my Replicator+ to constructing the bust with PLA plastic.

In about 3 hours, it had created a tiny replica of my face. My first impression was the the bust didn’t really look like me.

#deep-learning #ai #makers #3d-printing #deep learning