New Artificial Intelligence Renders Your Face in 3D Using Just a Photo
Until now, this has been a computational challenge of “extraordinary difficulty,” with obstacles posed by varied facial poses, expressions, and lighting.
Until now, this has been a computational challenge of “extraordinary difficulty,” the researchers wrote, with obstacles posed by different facial poses, expressions, and variable lighting.
“Typically if you want to reconstruct a face you have to try and use something called 3D morphable models, or shape from shading,” Aaron Jackson, a Ph.D. student who works on deep learning applied to human faces and who is one of the authors of the paper, told Seeker.
This method of 3D facial reconstruction analyzes shadows on the face to come up with a likely structure. It requires multiple images and poses.
“These techniques are really quite complicated,” Jackson remarked.
But AI, which can identify patterns that occur in sets of data and learn how to create its own, is becoming increasingly well positioned to tackle problems like this.
The team created a convolutional neural network, or CNN. Neural networks allow machines to learn — a bit like children do — by synthesizing the arrangement of interconnected neurons and synapses in the human brain.
Given enough data, neural networks can find patterns in any dataset, from faces to music to language, and reproduce them in a way that appears creative. The more data they are given, the more powerful and accurate the tool becomes.
Jackson and his colleagues — Adrian Bulat, Vasileios Argyriou, and Georgios Tzimiropoulos — fed their CNN tens of thousands of images of faces and their manually created corresponding 3D models.
Having used this data to identify how traits visible from a 2D image translate into the rest of a fully-realized face, the machine can now process an image of a face, make autonomous decisions about how that face probably looks in 3D, and create a corresponding model.
An interactive online demo of the technology allows visitors to see how it fared with the faces of Barack Obama and Marie Curie, among others. You can even upload your own photos, though the technology is limited only to human faces, for now. It’s really quite accurate, although can run into difficulties with facial hair and other features less likely to have cropped up during training.
Computer vision, as this area of research is known, has many potential applications.
“One of the most likely uses is in online shops where you buy glasses, so that you can try them on,” Jackson said. “There are also possible medical applications; perhaps you could simulate the results of plastic surgery.”
“Some other applications might be in facial expression analysis, measuring emotional arousing, for example, often used in psychological studies,” he added.
There’s clearly potential for computer games and augmented reality, too. You could potentially use it to render an accurate three-dimensional avatar of yourself in an online virtual reality game with friends, or use the technology to make human simulator games like The Sims more realistic.
Others have pointed out the potential for nefarious uses. Apple is replacing its TouchID fingerprint scanner with a 3D facial scan that can unlock your phone by identifying the owner’s face. Might a 3D rendering of somebody’s face be capable of fooling that software?
It wouldn’t be the first time that high-tech smartphone security features have failed: when Samsung’s Galaxy S8 smartphone launched with iris scanning, the sensor could be tricked with a color photo of the user’s eyes and the addition of a contact lens to make it appear real.
Jackson thinks it unlikely that the new tool can hack the latest iPhone.
“There is very little chance of it working on iPhones given that, apparently, Apple tested it on very high quality 3D models,” he said.
Jackson did concede that the technology could be used to improve surveillance techniques, given that recent research showed that facial recognition can be improved by having a 3D model instead of a single image.
Research into a deep learning facial recognition method that will be able to identify people even when they are wearing masks caused something of a furor recently. Zeynep Tufekci, a prominent academic who writes about tech and society, posted about its potential boon to authoritarian ends.
“The trend is clear,” she wrote. “Ever increasing new capability that will serve authoritarians well.”
“I suppose it could be used for things like that,” Jackson acknowledged. “But I wouldn’t like to think that it would be.”
WATCH: What Happens If We Give AI the Ability to Remember Everything?