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

Combination of Multiple Global Descriptors for Image Retrieval #313

Open
chullhwan-song opened this issue Feb 26, 2020 · 1 comment
Open

Comments

@chullhwan-song
Copy link
Owner

@chullhwan-song
Copy link
Owner Author

chullhwan-song commented Mar 5, 2020

image

  • Clova Vision, NAVER Corp 논문
  • mulit task = metric + image classification
    • image clasification과 learning deep feature를 동시에 가능하도록 설계
      • 참고로, 우리는 초기에 image clasification task에서 분류와 learning deep feature를 사용하였음. 하지만 초기의 연구들은 fc를 주로 사용했지만, 여기는 다양한 Pooling 기법을 적용하여 metric 학습.
    • 이전에 이런 비슷한 연구는 있었다. Fashion Style in 128 Floats: Joint Ranking and Classification using Weak Data for Feature Extraction #322
    • global descriptor (GD)는 실제 sift류의 local descriptor와 대비하여 사용하는 건데..
      • fc나 여기서 설명하는 feature는 GD가 맞긴하다.
      • 여기서 말하는 GD는 fc, SPoc, MAC등의 지금까지 이 git에서 많이 언급되었던 feature를 의미.
    • classification task > 우리가 그동안 전통적인 방식을 의미. 이때, GD도 fc인듯
    • metric task > 다양한 fc포함 GD 생성 > 이들을 concat하여 metric(여기서는 ranking loss라 표현)
      • concat할때, 단독이나..두 개이상.. S, M, G, SM, MS, SG, GS, MG, GM, SMG, MSG, GSM.
        • S:SPoc, G:GeM, M:Mac
        • SMG, MSG, GSM > 이 구성은 같은것 아닌가?
    • joint하게 두 개의 loss를 어떻게 섞어는지는 안나온듯~(간단히 보면..ㅠ)
      • 개인적으로 두 개의 loss를 + 연산으로만 해보니 궁금점이 있다. = triplet loss 와 cross entropy
        • 각각의 loss value의 scale이 다르다.
        • loss의 경사가 다르다.
          • 내 경우엔 cls만 하면 급격히 떨어졌고, 반면에 metric loss의 경우 아주 완면하게 떨어지는 현상이 각각 보였다. (각각의 학습할때만 보면,,)
        • 즉, 서로 좀 이질적인 성질을 가졌다 따라서, 두 개의 adapted weighted한 방법이 있지 않을까하는데..찾아보니..각각 가중치를 더 주는..이건 hyper parameter ㅎ
        • 합쳐서+해보니, 두 성질?(scale과 각각의 loss의 경사도)에 의해, 합의점이 와닿지 않았다. 하지만, 학습은 되는 듯하는데, 각각 할때의 성질과 비교해볼 때, 중간이 아니라 한쪽에 치우치는 loss 성격이 나오는듯..
        • 일단, 이 논문에서 두 loss의 접점을 어떻게 찾아가는 방법은 없었다. > 따라서, 그냥 + 한것같다.
  • backbone nets > ShuffleNet-v2 [35], ResNet-50 [17], SEResNet-50 [20] with ImageNet ILSVRC-2012
    • 특이점은 논문에, down-sampling operation between Stage 3 and Stage 4 & 그림 참조 > 왜 제거 되었는지 이유가 없다. ㅠ
  • label smoothing (LS), temperature scaling (TS) 사용 > 공부해야할 듯..(어디서 본 것 같은데.자세한.기억이.ㅠ)

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