“DeepFace: Closing the Gap to Human-Level Performance in Face Verification,” by Yaniv Taigman, Ming Yang, Marc’Aurelio Ranzato, and Lior Wolf of Facebook AI Research, was presented at the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) in 2014. It marked the moment deep learning overtook the older eigenface and hand-engineered approaches to face recognition.
DeepFace combined explicit 3D face modeling for alignment - warping each face to a frontal pose before analysis - with a deep neural network containing more than 120 million parameters across nine layers. The network was trained on a private Facebook dataset of four million facial images belonging to more than 4,000 identities, an order of magnitude larger than the public datasets of the time.
On the standard Labeled Faces in the Wild (LFW) benchmark, DeepFace reached 97.35% accuracy, reducing the error of the previous state of the art by more than 27% and coming close to the roughly 97.5% accuracy of humans on the same task. For a problem that had resisted decades of incremental progress, closing nearly all of the remaining gap in a single paper was a striking result.
DeepFace is also notable for what it represented institutionally: a major social-media company, sitting on billions of tagged photos, applying deep learning to faces at scale. Within a year Google’s FaceNet would push LFW accuracy past 99%, and face recognition would move from a research curiosity into widely deployed - and widely contested - commercial and government systems.