Taking pictures of yourself only to end up with your eyes closed in the shot is an all too common scenario for a lot of people. In fact, if you're in a social group of any kind, you're likely to have that friend who just can't keep their eyes open for the photo, no matter how many times you try again.
Facebook seems to be looking to use AI and deep learning to fix those pictures for you, and while the technology itself isn't all that new, the company is leveraging its massive pool of data to make it work better than other systems out there.
Typical deep neural networks already try to change a person's closed eyes to open ones, but, according to Facebook's research paper, there's no guarantee that the "new" eyes will be anything close to those of the person in the picture. That's because they're simply picking sets of eyes from people with a similar facial structure, whose eyes may be very different.
Instead, Facebook is using generative adversarial networks (GANs), a technology which has been used, for instance, by Microsoft to generate images from scratch based on text input from users, but the social network is using it to add information to existing pictures.
Using conditional and exemplar GANs, the company can not only create limitations for which information should be added to picture, but it can also ensure that the additional information is directly related to the subject of the photo, so that instead of adding a random pair of eyes to a picture, it'll use the eyes of the actual person in the picture. With Facebook already having access to so many of users' photos, it's bound to find more relevant information easier than other similar tools.
For now, this is only a research paper, but it shows that the company is interested in the technology, and it could very well become a part of Facebook's offerings sometime in the future.
Source: Facebook (research paper) via CNET
15 Comments - Add comment