It’s a terrible curse. The flash goes off, and you blink, resulting in an image that makes you look ridiculous. Ugh! But Facebook thinks it doesn’t has to be that way. Yes, they have developed new tool for that friend who never seems to have his eyes open during photos. And, I can’t say this enough, but 2018 has been very promising!
The tool uses artificial intelligence to retouch the eyes. The working is much like Adobe Photoshop’s “Content Aware” tool, which takes a smart guess at an item’s surroundings and fills in what it thinks should go there when photographer’s attempt to remove certain items from a photo: like fences, power lines, or even other people.
Facebook’s approach operates on much the same idea, but uses a Generative Adversarial Network (GAN) that teaches two separate machine learning systems. One system is trained to recognise whether images of open eyes are real, and the other exists to try and confuse it. In simple words, the latter system exists to build the most convincing images possible in an effort to fool the former.
However, the tool is still in its testing phase. And according to TechCrunch, the results are quite convincing, fooling testers more often than not into believing the computer-generated eyes were real — or at least getting them to admit they weren’t sure.
Also, like everything, the tool isn’t perfect just yet! Apparently, the main difficulties are with colour matching, that is required to blend skin tones in some photos. The eyelid, for example, contains millions of colours depending skin tone, light placement, and a tonne of other factors!
The research paper Facebook published about its new tool explains that, “given a training set of sufficient size, the network will learn what a human face ‘should’ look like, and will in-paint accordingly.” But most in-painting techniques have been unable to preserve a specific subject’s identity, which can lead to “undesirable and biased” (by which they presumably mean creepy, unrealistic, or inaccurate) results.
There is still no information as to when the new feature will be rolled out, but stay tuned for more updates on this!