Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

FaceNet: A Unified Embedding for Face Recognition and Clustering #43

Open
chullhwan-song opened this issue Aug 14, 2018 · 1 comment
Open

Comments

@chullhwan-song
Copy link
Owner

https://arxiv.org/abs/1503.03832

@chullhwan-song
Copy link
Owner Author

chullhwan-song commented Aug 14, 2018

what ?

  • triplet loss 기반 face embedding feature > 128차원 > facenet
    image
  • face verification
  • triplet selection - semi hard negative sampling

face net

image

  • Triplet Loss
    image
    • anchor와 pos는 가까워지도록, anchor와 neg는 멀어지도록 학습한다. 이를 위해,
      image
      • α 는 margin
    • loss
      image
  • triplet selection
    • 샘플자체가 너무 anchor와 pos는 너무 가깝고, anchor와 neg는 너무 멀면 학습에 그리 도움이 되지 않는다. 그리고 이런것까지 학습시키기는 너무 오래걸리고 많다.
    • 그래서, How?
      • anchor와 pos는 멀고, anchor와 neg는 가까운 sample을 뽑아보자는게 목적.
      • 그러나, 모든 sample에 대해 선택하는것은 불가능 > 계산량이 어마어마할것이다.
      • 그래서, batch 단위로 계산한다.
        • Offline: n step 마다 모델의 checkpoint를 생성하고 일부 데이터에 대해 뽑기
        • Online: mini-batch에서 뽑기
      • online 기법 이용
    • mini-batch 단위에서
      • semi-hard negative sampling
        image
      • facenet open source에서는 vgg face sampling(hard negative sampling)도 볼수 있음.

cnn

image
image

실험

  • 학습데이터
    image
  • Performance on LFW
    • fixed center crop: 98.87%±0.15
    • face alignment: 99.63%±0.09
      image
  • dimension
    image

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Projects
None yet
Development

No branches or pull requests

1 participant