issues
search
chullhwan-song
/
Reading-Paper
151
stars
26
forks
source link
FaceNet: A Unified Embedding for Face Recognition and Clustering
#43
Open
chullhwan-song
opened
6 years ago
chullhwan-song
commented
6 years ago
https://arxiv.org/abs/1503.03832
chullhwan-song
commented
6 years ago
what ?
triplet loss 기반 face embedding feature > 128차원 > facenet
face verification
triplet selection - semi hard negative sampling
face net
Triplet Loss
anchor와 pos는 가까워지도록, anchor와 neg는 멀어지도록 학습한다. 이를 위해,
α 는 margin
loss
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
facenet open source에서는 vgg face sampling(hard negative sampling)도 볼수 있음.
cnn
실험
학습데이터
Performance on LFW
fixed center crop: 98.87%±0.15
face alignment: 99.63%±0.09
dimension
https://arxiv.org/abs/1503.03832