endywon / styleeval

This is the official implementation of paper "Evaluate and Improve the Quality of Neural Style Transfer" (CVIU 2021))

StyleEval

[update 1/14/2022]

This is the official implementation of paper "Evaluate and Improve the Quality of Neural Style Transfer" (CVIU 2021)

Environment Implemented:

• Python 3.6
• TensorFlow 1.4.0
• Numpy 1.16.0
• CUDA 8.0

Getting Started:

Step 1: clone this repo

git clone https://github.com/EndyWon/StyleEval
cd StyleEval

Step 2: download pre-trained vgg19 model

bash download_vgg19.sh

Quality Evaluation:

• Test on a single image pair

python quality_criteria.py --content ContentImagePath --style StyleImagePath --stylized StylizedImagePath

MO-Net:

python MONet.py --content inputs/Seth.jpg --style inputs/Gogh.jpg

Discussions

For quality evaluation, we tested images of size 512x512 on an Intel Core i7-5820K CPU @ 3.30GHz × 12, the average time is about 126 seconds. We found that the most time was spent on calculating LP scores (about 80 seconds).

Currently the evaluation time is too slow, one may consider using GPU convolution to accelerate the patch matching process in LP calculation.

Citation:

If you find this code useful for your research, please cite the paper:

@article{wang2021evaluate,
  title={Evaluate and improve the quality of neural style transfer},
  author={Wang, Zhizhong and Zhao, Lei and Chen, Haibo and Zuo, Zhiwen and Li, Ailin and Xing, Wei and Lu, Dongming},
  journal={Computer Vision and Image Understanding},
  volume={207},
  pages={103203},
  year={2021},
  publisher={Elsevier}
}

styleeval's Issues

CST and MO-Net

Hi,
thank you very much for sharing the code for this interesting paper.
as far as I understand this is just the evaluation method, so basically the first contribution of the paper but not the second one with the CST and MO-Net right?

If thats the case, are you also going to publish the code for CST and MO-Net?

Thank you for clarifying.
Best regards

关于去除LP之后得到的值与使用完整指标不一致的问题

您好，感谢您提供这项工作的代码，但是我的电脑配置太低，于是我去除了代码中的LP部分，但是得到的值与之前使用完整的指标不一致，希望能够得到您的回复，谢谢！