Paper Review 15: DeepFace: Closing the Gap to Human-Level Performance in Face Verification

Summary

• Use ConvNets to extract features and calculate distance using them.
• Training method
  1. Train ConvNet first to extract feature
  2. Then, train Verification using X^2 or Siamese
• Used 3D modeling for face alignment before using ConvNet

Face Alignment (Frontalization)

Warp a detected facial crop to a 3D frontal mode using fiducial points

Training Method

1. Train ConvNet first to extract features
   1. Train like traditional AlexNet (Learning to maximize the prob of correct class)
2. Apply Verification Metric (learning unsupervised metric for generalization)
   1. Weighted X^2 distance

   2. Siamese Network