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

Photo Aesthetics Ranking Network with Attributes and Content Adaptation #65

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

Comments

@chullhwan-song
Copy link
Owner

http://users.eecs.northwestern.edu/~xsh835/assets/eccv2016_aesthetics.pdf

@chullhwan-song
Copy link
Owner Author

chullhwan-song commented Jan 23, 2019

소개

  • 이미지에 대한 Aesthetics 관점에서의 Ranking
  • 단순히 말하면 Image Quality에 대한 내용
  • we propose to learn a deep convolutional neural network to rank photo aesthetics in which the relative ranking of photo aesthetics are directly modeled in the loss function.
    • 이미지의 Aesthetics ranking을 위해 CNN을 이용하고 관련된 loss등을 적용하여 이를 성취.
  • Aesthetics with Attributes Database(AABB) 데이터셋 <- release
    • 다수의 익명의 사람이 aesthetic score와 의미있는 Attribute를 annotator한 데이터셋.
  • 이를 이용하여.. Attribute관점에서 Aesthetics score를 를 효과적으로 .. 다음 그림을 보면,,
    image
  • 제안한 CNN 모델을 통해 SOTA 달성.
  • 새로운 sampling 전략 for 학습

AABB 데이터셋

image

  • 11개의 attributes
    • interesting content, object emphasis, good lighting, color harmony, vivid color, shallow depth of f ield, motion blur, rule of thirds, balancing element, repetition, and symmetry.
  • total 10000개 이미지
    • we randomly split the dataset into validation (500), testing (1,000) and training sets (the rest).
  • Aggregating multiple raters 허용, 이를 "average ratings are well fit by a Gaussian distribution."
    • AVA 데이터셋은 아니다..

Fusing Attributes and Content for Aesthetics Ranking

  • fine-tuning AlexNet
  • fine-tune a Siamese network
    • image pairs as input and is trained with a joint Euclidean and ranking loss

Regression Network for Aesthetics Rating

  • fine-tuned from AlexNet
  • softmax loss가 아닌 Euclidean loss 로 대체
    image
    • y_i는 라벨된 데이터(GT)로, " the average ground-truth rating for image_i"
    • j^_i는 당연히 예측결과를 의미
    • [0,1] 사이의 값으로 scaling

Pairwise Training and Sampling Strategies

  • Euclidean loss에 대한 보완할 필요가 존재. > 비슷한 평균 Aesthetics 값을 가진 이미지들에서는 문제 발생소지가 있음.
  • 이를 위해, Siamese network
    • pairwise ranking loss to explicitly exploit relative rankings of image pairs available in the AADB data
      image
  • ranking loss > 알파는 margin과 관련된 param
    image
    image
  • jointly loss - Fig.3 a)
    image

Attribute-Adaptive Model

  • Attribute속성과 관련하여 aesthetic 에 반영하려는 개념.
  • Fig.3 b)
  • Attribute에 대한 분류 layer를 따로 두고 이 값과, Aesthetics 값과 concat
    • The attribute predictions from this layer are concatenated with the base model to predict the final aesthetic score. > 근데 그림상에서는 두 value간의 sigmoid activation 통과 한후, concat
      • 0~1사이로 rescale성??
        image

Content-Adaptive Model

  • Content 에 대한 결합. 요 개념은 밑의 그림보면
    image
  • Fig3. c)
  • 그냥 카테고리 분류같다.(실제적으로 AABB 데이터를 봐야할듯한데,, 전경이나 사람나오는 이미지등으로 분류가 되지 않았을까?)
  • We fine-tune the top two layers of AlexNet with softmax loss to train a content specific branch to
    predict category labels
  • 결론적으로 이미지가 좋냐나쁘냐와 각 attributes(선명함/조명등등) 마지막으로 각 content(카테고리)에 대한 조합하여 최종적인 aesthetic score를 계산하고픈것 같다.

성능

  • 자기네들이 첨 배포했는데 가장좋을듯..ㅎ
    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