How to compute FVD on UCF101

[MajorDavidZhang]

Hi, I'd like to compare the generation quality of my model with yours on UCF101 using the FVD score. After reading the papers and all the related issues in your repository, I still have some confusion about how to compute FVD on UCF101.

Since UCF101 has 101 classes, I plan to generate 20 clips per class, resulting in a total of 2020 clips. Following that, I intend to uniformly sample 4 clips per class from the UCF101 dataset and compute the FVD score between the generated clips and the real dataset clips using calc_metrics_for_dataset.py.

Is the above procedure correct?

Additionally, do I need to align the class information when organizing the generated and real datasets? For example, if Clips 0-3 in the generated data belong to UCF101 Class 1, should Clips 0-3 in the sampled real data also come from UCF101 Class 1?

Thank you for your time and response.