EdgeFace hits leaderboards with lightweight hybrid face recognition model

EdgeFace, a lightweight face recognition model optimized for mobile or edge device applications, is now state-of-the-art for all benchmarks on the Papers with Code leaderboard.

Unlike “heavier” models, EdgeFace does not rely on deep neural networks for computer vision, making it ideal for integration into edge-based devices that have limited computational resources. It is state of the art across datasets including LFW, Age DB-30, CFP-FP, IJB-B, IBJ-C, CALFW and CPLFW.

Professor Sebastian Marcel posted the news to LinkedIn, adding that the same team’s SynthDistill technique to train a lightweight “TinyFaR face recognition model” by “distilling the knowledge of a pre-trained teacher face recognition model using synthetic data” is state-of-the-art on a separate Papers with Code leaderboard for lightweight face recognition.

The team responsible is the Biometrics Security & Privacy (BSP) group at Idiap Research Institute. Their EdgeFace paper was published in IEEE Transactions on Biometrics, Behavior, and Identity Science ( Volume: 6, Issue: 2, April 2024).

Per the abstract, Edgeface is a hybrid model that combines the strengths of both convolutional neural networks (CNN) and Transformer models, as well as a low rank linear layer, to achieve high accuracy in face recognition while maintaining low computational costs and compact storage.

It is based on the pre-existing EdgeNeXt architecture, which was modified for efficiency in data processing.

The code for EdgeFace is publicly available.

Article Topics

accuracy  |  biometric testing  |  biometrics  |  biometrics at the edge  |  EdgeFace  |  facial recognition  |  Idiap