ArtFID: Quantitative Evaluation of Neural Style Transfer
Matthias Wright and Björn Ommer.
The field of neural style transfer has experienced a surge of research exploring different avenues ranging from optimization-based approaches and feed-forward models to meta-learning methods. The developed techniques have not just progressed the field of style transfer, but also led to breakthroughs in other areas of computer vision, such as all of visual synthesis. However, whereas quantitative evaluation and benchmarking have become pillars of computer vision research, the reproducible, quantitative assessment of style transfer models is still lacking. Even in comparison to other fields of visual synthesis, where widely used metrics exist, the quantitative evaluation of style transfer is still lagging behind. To support the automatic comparison of different style transfer approaches and to study their respective strengths and weaknesses, the field would greatly benefit from a quantitative measurement of stylization performance. Therefore, we propose a method to complement the currently mostly qualitative evaluation schemes. We provide extensive evaluations and a large-scale user study to show that the proposed metric strongly coincides with human judgment.
> pip install art-fidCUDA_VISIBLE_DEVICES=0 python -m art_fid --style_images path/to/style-images --content_images path/to/content-images --stylized_images path/to/stylized-imagesThe content images and the corresponding stylized images are compared in pairs. In order to ensure that a content image is matched up with the correct stylized image, both the content images and the stylized images are processed in lexicographical order. A simple way of pairing the content images and the stylized images is to use the name of content image for the corresponding stylized image.
--batch_size - Batch size for computing activations.
--num_workers - Number of threads used for data loading.
--mode - Evaluate ArtFID or ArtFID_infinity, choices = ['art_fid', 'art_fid_inf'].
--content_metric - Content metric, choices = ['lpips', 'vgg', 'alexnet'].
--device - Device to use, choices = ['cuda', 'cpu'].
--style_images - Path to style images.
--content_images - Path to content images.
--stylized_images - Path to stylized images.
The dataset is contained in artfid_dataset.csv. It consists of 250k labeled artworks.
- The implementation of the FID is based on mseitzer/pytorch-fid.
- The implementation of the FID_infinity is taken from mchong6/FID_IS_infinity.
- The implementation of the Inception network is taken from pytorch/vision.
- The checkpoint is hosted on the Huggingface Model Hub.
@article{wright_gcpr_2022,
title={ArtFID: Quantitative Evaluation of Neural Style Transfer},
author={Matthias Wright and Bj{\"o}rn Ommer},
journal={GCPR},
year={2022}
}