Face Identification and Verification Under Computational and Security Constraints

Face verification methods have undergone significant changes through the last decade. Rapid increasing of hardware computing power allows engineers to create neural networks to solve face verification problems. Classical computer vision face verification such as Viola-Jones Algorithm or Haar cascades Algorithm were forced out by deep learning Siamese Networks approaches. Nowadays we are faced with the challenge of finding a balance between accuracy and performance. Many light-weighted “mobile” models have good computational performance but lower accuracy in unconstrained data. In this paper, we provide some ideas of modifying the MobileFaceNet approach to increase accuracy without falling the evaluation performance.