Awesome

(Fréchet Denoised Distance) FDD_pytorch

Official Implementation of our Paper "Enhancing Plausibility Evaluation for Generated Designs with Denoising Autoencoder".

:technologist: To calculate the FDD score between two sets of images:

from fdd_tool import calculate_fdd

# set_1_images and set_2_images contain respectively original data and generated data (shape of (N, H, W, C)).

fdd_score = calculate_fdd(set_1_images, set_2_images)
print('the Frechet denoised distance is:', fdd_score)

:file_folder: Dataset

Make sure to import and save the dataset under the folder Data/

• The BIKED dataset is accessible via : https://decode.mit.edu/projects/biked/
• The ImageNet dataset is accessible via : https://www.image-net.org/download.php
• The FFHQ dataset can be obtained from: https://github.com/NVlabs/ffhq-dataset
• Seeing3DChairs

:test_tube: Sensitivity Test

See the main paper We provide the implementation of the levels of various disturbances, together with the distance metrics FID, FDD, TD and FD_Dino

:link: Cite The Paper

If you find our work or code helpful, or your research benefits from this repo, please cite our paper:

@article{fan2024enhancing,
  title={Enhancing Plausibility Evaluation for Generated Designs with Denoising Autoencoder},
  author={Fan, Jiajie and Trigui, Amal and B{\"a}ck, Thomas and Wang, Hao},
  journal={arXiv preprint arXiv:2403.05352},
  year={2024}
}