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

Composition-preserving Deep Photo Aesthetics Assessment #66

Open
chullhwan-song opened this issue Jan 22, 2019 · 1 comment
Open

Composition-preserving Deep Photo Aesthetics Assessment #66

chullhwan-song opened this issue Jan 22, 2019 · 1 comment

Comments

@chullhwan-song
Copy link
Owner

https://www.cv-foundation.org/openaccess/content_cvpr_2016/papers/Mai_Composition-Preserving_Deep_Photo_CVPR_2016_paper.pdf

@chullhwan-song
Copy link
Owner Author

Abstract

  • 최근 연구는 CNN기반으로..어느정도의 성과.
  • 그러나, 고정된 input size(fixed-size input)로 인해 문제가 발생.
    • fully connected layer의 사용함으로, 고정된 input size를 받는다..> SPPNet이 이런 문제를 해결하려고..
  • 입력이미지 damages > Image Composition에 대해 문제를 발생
    • transformed via cropping, scaling, or padding
    • reduces image resolution, or causes image distortion
      image
  • 이 연구에서는 이미지의 Composition를 보존하는 방법 제시 - composition-preserving deep ConvNet method
    • directly learns aesthetics features from the original input images without any image transformations.
    • 이 연구에서는 오리지널 사이즈를 그대로 받음. 이를 위해 "adaptive spatial pooling layer" adding
  • 결론적으로 "Multi-Net Adaptive Spatial Pooling ConvNet architecture" 란걸 제안.

Composition-preserving Deep Network for Photo Aesthetics Assessment

  • 역쉬 SPPNet에서 영감을 얻음.
    • 하지만 이와는 달리 " multiple fixed-size input" 를 받는 형태
    • multiple sub-networks for different pooling sizes > ??

Composition-preserving Deep ConvNet

  • 기본적으로 언급했던, SPPNet의 adaptive spatial pooling 전략을 취함. > 이는 오리지널 입력이미지의 크기 그대로 받기 위함.
    image
    image

Multi-Net Adaptive-Pooling ConvNet

  • SPPNet의 기본 개념에서 multi-scale feature 추출하는게 목표.
    • SPPNet넷의 위의 그림처럼 multi-sclae하지 않나??? 다양한 Grid개념...
  • 이 연구에서는 SPPNet자체가 하나의 conv layer(맨 마지막 conv layer)를 입력으로 받는다는게 문제라고 봄.
  • 그래서, 다양한 사이즈의 pooling layer들을 결합함 이를 "Multi-Net Adaptive-Pooling method"
    • 맨마지막이 아니라 하위의 여러 layer에서 받는 case인듯..
  • 최종적으로 Multi-Net Adaptive-Pooling ConvNet (MNA-CNN) 제시
    • basenet에서의 sub-networks들을 받음(바로 위에서 언급한것이 맞는듯..) > SSD를 생각하면 쉬움.
    • 관련 그림이 그닥 와닿지는 않음.
      image

Scene-Aware Multi-Net Aggregation

  • 최종 결과는 multiple sub-network에서의 각 결과들을 average
  • 실제로는 위의 평균을 내는 루틴을 다음과 같은 것으로 대체
    • sub-network predictions 과 image scene categorization posteriors를 concat하는 형태로
      • a new aggregation layer that takes the concatenation of the sub-network predictions and the image scene categorization posteriors as input and output the final aesthetics prediction.
  • 다음 그림을 보면 명확
    image

Experiments

image
image
image
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