Triplet Loss

1. 개념

• 무작위 dataset을 뽑은 후, positive pair와 negative pair들과 비교
  • positive는 가까이, negative는 멀리 배치

 

2. 계산 방법

• $L(A, P, N)=\max(\Vert f(A)-f(P) \Vert^{2}-\Vert f(A)-f(N) \Vert^{2}+\alpha, 0)$
  • $A$는 anchor input, $P$는 positive pair, $N$은 negative pair, $f$는 embedding, $\alpha$는 positive pair와 negative pair 사이의 margin

 

참고 링크

https://mic97.tistory.com/16

https://soobarkbar.tistory.com/43