From a single picture, this neural network can make a 3D model of your face. That's a whole extra dimension for free!
Researchers at the University of Nottingham and Kingston University have wielded the powerful sword of Convolutional Neural Networks (CNNs) to slice through the Gordian knot of getting a 3D image from 2D.
Intuitively, it makes sense. We know enough about what faces are to be able to guess the shape of someone's head when we see a photo of them. There are clues from where shadows are being cast, reflections on a shiny forehead, et cetera. But this is one of those things that until recently seemed easy, but was was not easy.
You can probably come up with a few rules of thumb. Light parts are probably more protruding and more upward-facing than dark parts, the middle of the head sticks out more than the top and the bottom, and so forth. That's a very manual process of trying to identify and measure clues from the image and relate them to the output, but that's roughly what face reconstruction used to do (with a veneer of science over the top).
The great power of CNNs, and neural networks in general, is that you don't have to find all those "clues" yourself - the network gradually notices patterns and gets a sense of what's important as a natural consequence. The challenge is that it's extremely sensitive to how you train it.
The classic example, although it's probably not true, is about a group of scientists commissioned by the military to make an AI that can spot camouflaged tanks in a forest. They train it on images with camouflaged tanks, and images without them, and over time the AI becomes 95% accurate at spotting these tanks. They hand it over to the military to test out in the real world, but it ends up only guessing as good as random chance. What happened? The pictures of camouflaged tanks were taken on a cloudy day, and the pictures without them were taken on a clear day. The AI had just learnt to spot the weather.
So although, to some degree, this approach is as simple as throwing data at a neural network ... that doesn't give enough credit to how carefully the researchers have had to build up an enormous, high-quality and unbiased dataset.
Applications? Games / TV / movies, historical reconstruction, building a model of someone's head before an accident, and probably thousands of far better examples than those ones off the top of my head.
Or, y'know, selfies with depth.
hose who have sworn off 10-minute mental breaks that turn out to be 60-minute reveries had best avoid a fascinating new way to see how your face looks like in 3-D mode. A University of Nottingham and Kingston University team have actually come up with a way to turn a 2-D photo of a face into a 3-D model. A new algorithm "learned" how to make a 3-D model from a flat image. You can check out an online demo of their paper, thanks to the team. They said, "Please use a (close to) frontal image, or the face detector won't see you (dlib)."