extras

Video Evaluation with Fréchet Video Distance (FVD)

We provide a copy of the official implementation of FVD to facilitate the reproduction of our results with RELATE's temporal extension. Please cite the authors of FVD, if you use their code for video evaluation:

@article{unterthiner2018towards,
  title={Towards Accurate Generative Models of Video: A New Metric \& Challenges},
  author={Unterthiner, Thomas and van Steenkiste, Sjoerd and Kurach, Karol and Marinier, Raphael and Michalski, Marcin and Gelly, Sylvain},
  journal={arXiv preprint arXiv:1812.01717},
  year={2018}
}

To minimize compatibility issues with the remainder of the RELATE codebase, we provide a separate conda environment for the FVD computation fvd_environment.yml. Please build it separately using:

conda env create -f fvd_environment.yml
conda activate fvd

After the environment is set up, you can compute the FVD scores for two video directories using:

(fvd) user@host:~$ python compute_fvd.py path/to/sampled_videos_from_model path/to/test_videos_from_dataset

Please note that the videos need to be in GIF format in order to be recognized by the script!