SAMScore: A Semantic Structural Similarity Metric for Image Translation Evaluation

Yunxiang Li, Meixu Chen, Wenxuan Yang, Kai Wang, Jun Ma, Alan C. Bovik, You Zhang.

Quick start

Run pip install samscore.

pip install git+https://github.com/ChaoningZhang/MobileSAM.git
pip install timm

pip install samscore
pip install git+https://github.com/facebookresearch/segment-anything.git

The following Python code is all you need.

import requests
import os
import samscore

def download_image(url, save_path):
    response = requests.get(url)
    response.raise_for_status()  # Raise an exception if the request was unsuccessful

    with open(save_path, 'wb') as file:
        file.write(response.content)
os.makedirs('imgs', exist_ok=True)
# Example usage
image_url = 'https://i.ibb.co/yFFg5pn/n02381460-20-real.png'
save_location = 'imgs/real.png'
download_image(image_url, save_location)

image_url = 'https://i.ibb.co/GCQ2jQy/n02381460-20-fake.png'
save_location = 'imgs/fake.png'
download_image(image_url, save_location)

## Initializing the model
# You can choose "vit_t", "vit_l", "vit_b", "vit_h"
SAMScore_Evaluation = samscore.SAMScore(model_type = "vit_b" )
samscore_result = SAMScore_Evaluation.evaluation_from_path(source_image_path='imgs/real.png',  generated_image_path='imgs/fake.png')

print('SAMScore: %.4f'%samscore_result)

Use it on pytorch

import requests
import os
import samscore
import cv2
import torch

def download_image(url, save_path):
    response = requests.get(url)
    response.raise_for_status()  # Raise an exception if the request was unsuccessful

    with open(save_path, 'wb') as file:
        file.write(response.content)
os.makedirs('imgs', exist_ok=True)
# Example usage
image_url = 'https://i.ibb.co/yFFg5pn/n02381460-20-real.png'
save_location = 'imgs/real.png'
download_image(image_url, save_location)

image_url = 'https://i.ibb.co/GCQ2jQy/n02381460-20-fake.png'
save_location = 'imgs/fake.png'
download_image(image_url, save_location)


# You can choose "vit_t", "vit_l", "vit_b", "vit_h"
SAMScore_Evaluation = samscore.SAMScore(model_type = "vit_b" )

source_cv2 = cv2.imread('imgs/real.png')
source = torch.from_numpy(source_cv2.transpose(2, 0, 1)).unsqueeze(0).float()
source = torch.cat((source,source,source),dim=0)

generated_cv2 = cv2.imread('imgs/fake.png')
generated = torch.from_numpy(generated_cv2.transpose(2, 0, 1)).unsqueeze(0).float()
generated = torch.cat((generated,generated,generated),dim=0)

samscore_result = SAMScore_Evaluation.evaluation_from_torch(source,  generated)

print('SAMScore:',samscore_result)

Citation

If you find this repository useful for your research, please use the following.

@article{li2023samscore,
  title={SAMScore: A Semantic Structural Similarity Metric for Image Translation Evaluation},
  author={Li, Yunxiang and Chen, Meixu and Yang, Wenxuan and Wang, Kai and Ma, Jun and Bovik, Alan C and Zhang, You},
  journal={arXiv preprint arXiv:2305.15367},
  year={2023}
}