MANILA, Philippines - Have you ever spotted yourself with your eyes closed having been caught mid-blink in group photos posted on social media? Facebook’s new AI research may just have the eye-opening solution to this problem.
While the dawn of Photoshop has provided users with tools to enhance or edit out undesirable features, the complex details of human eyes have always been difficult to change realistically.
What Facebook does is use a Generative Adversarial Network (GAN), a machine learning system comprised of two neural networks that "contend" with each other. What this essentially means is that one part of the system recognizes faces while the other replicates the image based on what was recognized, drawing from a huge database of people's photos.
Facebook’s researchers uses countless photos of people with eyes both opened and closed uploaded in their platform as reference data for the GAN system. This allows the GAN to learn what eyes should go on what person including its shape, color, and other subtle details.
The results were reportedly realistic enough to fool the people they showed it to. Supposedly, there are no signs of stitching or color mismatch for the person’s skin. See for yourself:
Photo from Facebook
It still fails in some situations when a person’s eyes are partially covered by hair or other objects but, these problems are said to be fixable.
This tool once fully developed should make it easy for people to fix photos where they were snapped blinking, saving them the time and effort from using more complicated editing software programs. – Rappler.com