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

Faster R-CNN Features for Instance Search #37

Open
chullhwan-song opened this issue Aug 7, 2018 · 1 comment
Open

Faster R-CNN Features for Instance Search #37

chullhwan-song opened this issue Aug 7, 2018 · 1 comment

Comments

@chullhwan-song
Copy link
Owner

intro: DeepVision Workshop in CVPR 2016
homepage: http://imatge-upc.github.io/retrieval-2016-deepvision/
arxiv: http://arxiv.org/abs/1604.08893
github: https://github.com/imatge-upc/retrieval-2016-deepvision

@chullhwan-song
Copy link
Owner Author

전체 구조

image

  • Faster R-CNN 기반 feature extraction
    • 참고로 ROI 기반 feature 추출은 faster rcnn이 가장 좋다고 알려져 있음 - 이전 techtalk 발표자들도 이런 언급한적이 있음.
    • Faster R-CNN implements a region pooling layer that extracts the convolutional activations for
      each of the object proposals learned by the RPN.
    • 위의 그림과 같이, 이를 이용하여 2 type descriptor 추출
  • Image-wise pooling of activations (IPA)
    • for global descriptor
    • MAC과 거의 유사
    • sum pooling이냐 max pooling이냐? 이에 대해 비교 실험.
    • one can choose to ignore all layers in the network that operate with object proposals and extract features from the last convolutional layer. > RPN 이전 conv feature map를 이용한다는 의미.
  • Region-wise pooling of activations (RPA)
    • descriptor by aggregating the activations of that window in the RoI pooling layer
      • R-MAC과 거의 유사 즉, 모든 ROI 영역에 대한 feature들을 이용
    • 각각의 region에 대한 feature들을 > sum, max > 이는 실험결과에 의해
      image
    • global descriptor IPA는 sum이 낫고, region 단위에서는 max(R-MAC)이 낫는듯~

기타 실험

image

  • 참고) attention 기반의 논문은 아님
  • 다만, fine-tuning 기반의 학습 모델에서, 즉, 그 학습셋에 맞게 attention되는 현상을 보여주고 있음.
  • 위의 방법은 어떻게 heatmap을 표현했는지에 대한 방법은 나오지 않음.

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