CSNet-Pytorch

This is an unofficial Pytorch implementation of Composition Scoring Model in Camera View Adjustment Prediction for Improving Image Composition(2021).

Composition Scoring Model

Composition Scoring Model, which we named as CSNet, predicts image composition score(0 ~ 1).

Performance

There is no evaluation metric in original paper because this model was used only when training view adjustment prediction model.

We used scored crops dataset(GAICD, CPC) as test dataset and calculate the accuracy of judging which image has better composition.

Gap in Accuracy means a score gap of test image pairs. For example, when Gap >= 0.5, the score gap of test image pairs is greater than or equal to 0.5.

Image Perturbation	Data Augmentation	Accuracy(Gap>=0.5)	Acccuracy(Gap>=1.0)
Shifting, Zooming-out, Cropping, Rotation	Shift Borders, Zoom-out, Rotation	71.8%	76.2%
Shifting	Shift Borders	73.8%	78.3%

Dataset

• GAICD [Paper][Github]
• FCDB [Paper][Github]
• CPC [Paper][Homepage]
• Unsplash [Github]

Usage

git clone https://github.com/PROLCY/CSNet-Pytorch.git
cd CSNet-Pytorch && mkdir weight

pip install -r requirements.txt

Demo

Download pretrained model in the directory weight

Image Perturbation	Data Augmentation	Accuracy(Gap>=1.0)	Download
Shifting, Zooming-out, Cropping, Rotation	Shift Borders, Zoom-out, Rotation	76.2%	Link
Shifting	Shift Borders	78.3%	Link

python demo.py {image_dir_path}

You can check the composition score of images in terminal.

Result

Predicted composition scores are as follows.

If you are interested in this repository, please contact ckd248@naver.com