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

Deep Supervised Hashing with Triplet Labels #86

Open
chullhwan-song opened this issue Feb 19, 2019 · 1 comment
Open

Deep Supervised Hashing with Triplet Labels #86

chullhwan-song opened this issue Feb 19, 2019 · 1 comment

Comments

@chullhwan-song
Copy link
Owner

chullhwan-song commented Feb 19, 2019

https://arxiv.org/abs/1612.03900
https://pdfs.semanticscholar.org/9a0c/8fbc6f4f48f736e0f25006c3a9d8c7da83e5.pdf

@chullhwan-song
Copy link
Owner Author

chullhwan-song commented Feb 19, 2019

Abstract

  • Hashing 은 ANN Search에서 popular & powerful 방법
    • for large-scale image retrieval
  • 전통적인 방법에서는(DL이 아닌) feature vector를 만들고 난후에 hash화함.
  • 이 과정에서 feature vector와 hash code사이에 optimal한 학습으로 인해 생성된것이 아님
    • end to end가 필요하다는 의미도 포함한듯
  • DL 방법은 end to end통해 feature vector와 hash code사이에 학습을 통해 optimal hash code를 발생하는 연구임.
    • 기존 연구에서는 softmax, pairwise label를 통해..
  • 이전의 SOTA인 DPSH는 pairwise label(loss)를 통해 성취
  • 우리의 연구는 triplet loss(label)를 이용함.
  • 기본적으로 DPSH의 연구에서 영감을 얻어으니 거의 같고, triplet loss만 다른듯

Contributions

  • triplet label based
  • 새로운 hash code 평가
  • loss function
  • SOTA performance

Approach

image
image

  • triplet loss는 많은 paper에서 많이보았기 때문에 넘어감 (밑의 리뷰 참고)
  • 다만 여기서는 vector가 아닌 hash code 형태 즉, hamming distance에 의해 기본 distance를 계산.
    • hamming이 아니더래도, hash 를 vector로 생각하여 일반적인 distance(inner product)를 사용해도 무방하다고 생각함.
  • Hash Function
    • 정확히 어떤 알고리즘으로 사용했는지는 안나타는듯..
    • refer 정도이니 단순희 sign/or others activation정도를 했을것으로 추정 > 뒤에 sign 함수를 쓰는 내용이.
  • Loss Function
    image
    image
    image
    • U 는 real vector(fc vector)

Experiment

image
image
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