A Blog by Jonathan Low

 

Jun 24, 2018

The Reason Facebook Uses Artificial Intelligence To Open Eyes In Blinking Selfies

Anything to enhance the user perception of impact - and self. The question is what happens when Facebook and others start making such changes without permission? JL

James Vincent reports in the Verge:

Facebook will do whatever it takes to encourage you to share more content — even, it turns out, using artificial intelligence to salvage photographs marred by an accidental blink.
Facebook will do whatever it takes to encourage you to share more content — even, it turns out, using artificial intelligence to salvage photographs marred by an accidental blink.
In a research paper published this week, a pair of Facebook engineers describe a new method that uses machine learning to retouch closed eyes in photos. The method is just at the research stage at this point, but while there’s no guarantee Facebook will ever introduce such a feature into its website or apps, that does seem like a strong possibility. As the researchers note, there are already plenty of photo retouching tools that can remove red eyes or smooth out blemishes, so why isn’t there one that solves the problem of accidental blinks?
The system works by using a machine learning method called a generative adversarial network, or GAN. This is a technique that has proved particularly adept at generating imagery, and it has already been put to work creating fake celebrities, changing the weather in videos, and even designing clothes. Blinking is no challenge at all.




Facebook’s system trains on pictures of the target with their eyes open, then applies that data to photos when they are blinking.
Image: Facebook

For the system to successfully edit a photo, it has to first learn what a user’s eyes normally look like. This means training it on images of people when they’re not blinking. From this, the software gets an idea of their eye shape and color and can then fill that information in for blinking images. (This process is referred to as “in-painting.”) It’s not a perfect solution, and the AI struggles to edit glasses, long fringes, and faces at extreme angles. But when it works, it creates incredibly realistic fakes like the ones you see above.
AI tweaks like this are only going to become more common. Software companies such as Adobe and Pixelmator are working on a number of machine learning-powered features that allow anyone to edit images with little technical skill. In the not-too-distant future, artificial intelligence will make not only small changes but generate completely new photos and video.
This will obviously create some problems for society as it becomes easier for anyone to fake visual evidence. But it has short-term challenges, too — including for Facebook. After all, the company frequently touts its site as a home for authentic expression, so will it give people AI tools to change their appearance? And if so, where will it draw the line? Editing out a blink seems pretty harmless, but more substantial changes might be problematic. AI editing is just getting started.

0 comments:

Post a Comment