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

Cross-dimensional Weighting for Aggregated Deep Convolutional Features #51

Open
chullhwan-song opened this issue Sep 13, 2018 · 2 comments

Comments

@chullhwan-song
Copy link
Owner

@chullhwan-song
Copy link
Owner Author

  • 포멧은 다른데 같은 논문이 : 링크

@chullhwan-song
Copy link
Owner Author

what?

  • 기존 CNN convolution feature map을 가지고 attention concept을 넣어 convolution feature map을 강화하여 image feature로써 representation
  • 하지만, 이 논문은 초기 논문으로, learn-able deep feature는 아니다.
  • 즉, 학습 과정에서 나온것이 아니라, 학습이 끝난 imagenet으로 학습 시킨 alext, vgg,resnet등을 이용한것이다. 일종의 handcraft feature라 볼 수 있다.

crow

  • representation
    • vgg - pool_5 feature extraction
    • spatially-local pooling : Sum-pooling or max-pooling
      image
      • location wise - 알파
      • channel wise - 베타
    • Cross-dimensional Weighting
      • image , k는 channels
      • 위의 그림에서, channel-wise = image location wise = image
      • weighted feature tensor = image
        image
      • 최종적으로, weighted feature tensor
        • aggregated by sum-pooling per channel = image
          image
          * channel 기준으로, 즉 k dimension feature size
          • 이후에, norm > pca > norm
        • 이를 "Cross-dimensional Weighted or CroW features"
  • Response Aggregation for Spatial Weighting
    • 즉, 위의 location weight : image 구성하는 법
    • the matrix of aggregated responses from all channels per spatial location, which we compute by summing feature maps(C)
      image
    • After normalization and power-scaling we get aggregated spatial response map S(수식.4)
      image
      • image
      • a = 0.5, b =2 (근데 실 소스에서는 2로 setting??)
  • attention 결과 - 주의할점은 channel은 이렇게 표시 할수 없다.. only spatial attention
    image
  • Sparsity Sensitive Channel Weighting
    Sparsity Sensitive Channel Weighting
    • 위의 그림.1에서 베타를 의미한다.
    • to derive a channel weighting based on the sparsity of feature maps.
      • 채널에 대한 sparsity 속성을 크게 와닿지않음 > 위의 공간적 특성은 와닿는데..?? > 실제 feature로써 역할을 하니 sparsity하면 구별성이 강한 feature는 발현되고 그렇지 않는 것들은 0
    • 당연히 feature map의 channel 차원
    • 각 채널마다 0 보다 큰 기눈으로 sparsity를 계산(캐치)
      image
      image
    • channel sparsities are highly correlated for images of the same landmark and less correlated
      for images of different landmarks.
      • 여기서 의미하는 sparsity는 각 클래스 또는 각 landmark(여기서 사용하는 데이터셋) 의 주요 특징을 의미하는듯으로 보임.
    • It appears that the sparsity pattern of channels contains discriminative information.
    • 드물게 나타나는 feature는 중요한 요소일수 있음. 그래서 inverse document frequency의 유사한 개념으로 weight를 구함.
      image
      • K에 대한 정보가 없다 ㅠ
      • image

실험험

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