DeepFace: Closing the Gap to Human-Level Performance in Face Verification

DeepFace targets face verification, where the conventional pipeline consists of detect, align, represent, and classify. The authors revisit the alignment and representation stages: they employ explicit 3D face modeling to apply a piecewise affine transformation for alignment, and derive the face representation from a nine-layer deep neural network. This network involves more than 120 million parameters and uses several locally connected layers without weight sharing rather than standard convolutional layers, and it was trained on the largest facial dataset to date, four million labeled images belonging to more than 4,000 identities.

Coupling accurate model-based alignment with the large facial database, the learned representation generalizes remarkably well to faces in unconstrained environments even with a simple classifier. DeepFace reaches 97.35% accuracy on the Labeled Faces in the Wild (LFW) dataset, reducing the error of the previous state of the art by more than 27% and closely approaching human-level performance, marking a step toward closing the gap between machine and human face verification.