FaceNet

• FaceNet: a Unified Embedding for Face Recognition and Clustering
• mapping from face images to a compact Euclidean space where distances directly correspond to a measure of face similarity
• Optimize the Embedding itself
• FaceNet directly trains its output to be a compact 128-D Embedding using a Triplet Loss function
• Choosing which triplets to use turns out to be very important for achieving good performance
  • inspired by Curriculum Learning
  • online negative exemplar mining strategy which ensures consistently increasing difficulty of triplets as the network trains
  • also explore hard-positive mining techniques which encourage spherical clusters for the embeddings of a single person
• squared Lp Regularization L2 distance, in the Embedding space directly correspond to face similarity: faces of the same person have small distances and faces of distinct people have large distances
• face verification simply involves thresholding the distance between the two embeddings; recognition becomes a KNN classification problem
• Labeled Faces in the Wild
• Zeiler Fergus
• Inception
• Harmonic Embedding