Facebook’s DeepFace facial verification project “closing the gap to human-level performance”


It’s no secret that Facebook has been interested in the world of facial recognition for a while and the company’s current project along those lines, artificial intelligence software called DeepFace, is supposed to be closing in on human-level performance in facial verification.

The MIT Technology Review reports that the project, which is slated to remain a research project for the moment, has seen a “significant advance” over previous facial recognition efforts. DeepFace is almost matching human recognition percentages, with a success rate of 97.25% irrespective of lighting or the direction that a person is facing. Provided the face is in view, of course. Humans in the same tests have attained a success rate of 97.53%.

Facial verification isn’t the same as recognition, verification entails recognising that two images are of the same face without worrying about identifying who that person actually is. DeepFace functions by turning a side-on image into a front-facing one by using a map of an “average” face. A simulated neural network then analyses the image looking for similarities to other faces, throwing up a match if enough are found.

Facebook has used more than 4 million images of around 4,000 different people (4.4 million and 4,030 respectively, according to The Verge) to train the DeepFace AI and these images have come from the collection that Facebook has at their disposal, so you might have been involved in improving the AI’s verification performance without even knowing it.

Source: The Verge


About Author

Leave A Reply