Usage of UCF-101 dataset

[snap-research]

Usage of UCF-101 dataset

[bluer555]

Hi,

we use all 13,320 videos for training without the split.

Thanks!

[sihyun-yu]

Thank you for the clarification!

[sihyun-yu]

Hello! I have one more question.
You compared with baselines which mostly uses "train" split for the generator for the evaluation. However, you use all data for training; isn't it unfair comparison?

Sincerely,
Sihyun.

[bluer555]

Hi Sihyun,

We follow the settings of DVD-GAN, the strongest baseline prior to our work.

The main drawback of the UCF-101 dataset is the relatively poor image generation quality (>40 FID), our video synthesis quality (FVD) is highly correlated to the image generation when the FID is not good enough. To be honest, we are not sure if using the training split can improve the FID or not. It is interesting to perform an experiment on this.

Thanks!