Facebook has recently developed a software program capable of facial verification called DeepFace.
DeepFace scored 97.25 percent on a challenge showing two faces of the same person. A human being also got the correct score and scored 97.23 percent. According to the researchers, this is a great achievement compared to past software programs of the same kind and DeepFace showed the potential that deep learning has. Deep learning is a branch of Artificial Intelligence that utilizes a large network of neurons to identify patterns in a massive amount of data.
"You normally don't see that sort of improvement," Yaniv Taigman, part of Facebook's AI Team told MIT News. "We closely approach human performance." He cited that the error rate has been reduced by more than a quarter relative to earlier software that can take on the same task.
The researchers stated that the new software does not perform facial recognition (identifying the face), instead, it is capable of facial verification or identifying two images that depicts the same face. This new capability will increase Facebook's accuracy in telling users who they should tag every time they upload new photos.
DeepFace uses two steps in order to successfully verify the image. First, it will revise the face's angle to make it look forward. When DeepFace does this, it's guided by a 3D model of what a "forward-facing face" looks like. After that, the software will use deep learning to come up with a numerical description of the re-angled face. Once DeepFace identifies similar descriptions of the face from two angles, it will confirm that the two images depict the same face.
Although DeepFace is showing promising results, it remains a research project for the meantime. Facebook published a research paper about DeepFace last week and the team working on it will talk about the new software on the IEEE Conference on Computer Vision and Pattern Recognition this coming June.
