Hi Junyao, sorry to interrupt. I have the following problems regarding this repository, I could be great if you can help me answer the problems.
I am still confused why do we make sure that the number of the videos and the length of each video should be the same in the calculation of FVD. Though this problem has been answered in issue 2 and issue 8.
In the original FVD paper, it does not specify that we should use the same number of videos and videos of the same length in the calculation of FVD. As they use the following equation to calculate.
Also in the code of StyleGAN, in their frechet_video_distance.py they calculate the FVD using the mean and covariance of the feature vectors.
I wonder if it is a common practice to evaluate generated videos against real videos with the same number of videos and the same video length?
I wonder if you have check the result of your implementation of FVD with the result of the original StyleGAN FVD and VideoGPT FVD?
Hi Junyao, sorry to interrupt. I have the following problems regarding this repository, I could be great if you can help me answer the problems.
I am still confused why do we make sure that the number of the videos and the length of each video should be the same in the calculation of FVD. Though this problem has been answered in issue 2 and issue 8. In the original FVD paper, it does not specify that we should use the same number of videos and videos of the same length in the calculation of FVD. As they use the following equation to calculate. Also in the code of StyleGAN, in their
frechet_video_distance.py
they calculate the FVD using the mean and covariance of the feature vectors. I wonder if it is a common practice to evaluate generated videos against real videos with the same number of videos and the same video length?I wonder if you have check the result of your implementation of FVD with the result of the original StyleGAN FVD and VideoGPT FVD?
Look forward to your reply!