MS-SWD

This is the repository of paper Multiscale Sliced Wasserstein Distances as Perceptual Color Difference Measures, which has been accepted by ECCV 2024.

---

Requirement

• Python>=3.7
• Pytorch>=1.8

Useage

from MS-SWD import MS_SWD

model = MS_SWD(num_scale=5, num_proj=128)
# X: (N,C,H,W)
# Y: (N,C,H,W)
distance = model(X, Y)

or

git clone https://github.com/real-hjq/MS-SWD
cd MS-SWD

python MS_SWD.py --img1 <img1_path> --img2 <img2_path>

News

The learned version of MS-SWD is available on IQA-PyTorch.

import pyiqa
import torch

device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
cd_measure = pyiqa.create_metric('msswd', device=device)

Citation

@inproceedings{he2024ms-swd,
  title={Multiscale Sliced {Wasserstein} Distances as Perceptual Color Difference Measures},
  author={He, Jiaqi and Wang, Zhihua and Wang, Leon and Liu, Tsein-I and Fang, Yuming and Sun, Qilin and Ma, Kede},
  booktitle={European Conference on Computer Vision},
  pages={1--18},
  year={2024},
  url={http://arxiv.org/abs/2407.10181}
}

Acknowledgements

Part of the code is borrowed from GPDM, and srgb2lab comes from flip_loss.py in ꟻLIP. Sincerely thank them for their wonderful works.