Behold the glorious future of neural networks: disembodied faces rotating in the darkness. Research submitted to Cornell University uses deep neural networks to create detailed 3D models of faces using a single 2D picture.
The results are remarkably accurate.
Researchers Hao Li, Shunsuke Saito, Lingyu Wei, Koki Nagano, and Liwen Hu wanted to create super detailed face models without the need for professional lighting or even a full photo. Face mapping at this level usually requires a series of photos in ideal lighting to make sure you get all the curves, angles, and asymmetries of the face. The researchers relied on an extensive “face database” (a real thing that exists) to make smart inferences on the finer details of each face. Neural networks create the face by filtering through a network of possible textures before scanning and then blending the pertinent facial features and skin tones.
In their submission, researchers hypothesise creating a fully realised 3D avatar in virtual reality and online gaming. You can already scan and input your face into games like NBA 2K17 (to mixed results), so get ready for a lot of photorealistic Muhammad Ali clones when you ball online. Or Alibaba CEO Jack Ma, if that’s your thing:
As with everything, there are privacy concerns for people who want to be arseholes. There are photos of everyone online, so what’s to stop someone from wearing a different face while trolling online? Imagine playing a 3D game and seeing yourself as the face of the avatar going Leroy Jenkins and ruining things for the team or, worse case scenario, harassing other players.
For now, though, let’s hope that the technology will be used as it’s intended, easy, high degree realism in virtual space. Maybe we can one day print it out URME style. Pioneered by “Leo Selvaggio” (a Bond-sounding pseudonym) URME are 3D printed masks that disguise you to face recognition software. So in sense, we could finally fulfil our dystopian dreams of a virtual world in which everyone is a photorealistic, 80's era Trump.
Gif credit: Prosthetic Knowledge
[h/t Carl Franzen]