martinetoering / vicc Goto Github PK

Hi, the UCF101 frame dataset I create according to the code you release is about 94G, but the data provided by CoCLR is about 28G. I want to know what lead to this difference.

Hi，I also have a question about pretraining. In single pretraining, did you train the model for 500 epoch but only use the checkpoint corresponding to the 299 epoch?

Hi,
In the pretrained model part, I can only see links for pretrained S3D model. Could you also provide pretrained R(2+1)D model? Thanks.

Hi,

Thanks for sharing your work and your code.

In your pretraining code, the transformations that crop, flip, blur, and jitter have the argument consistent = False. However, in your paper, you mentioned that:

Random cropping, horizontal flipping, Gaussian blur, and color jittering are used in a frame-consistent manner on RGB and flow clips following recent works [12,29].

If I understood correctly your code and launch it, the augmentations would be different for each frame in the clip, but I comprehended that the augmentations should be the same for the whole clip from your paper.

Could you provide me with some guidance on this issue, please?

Thank you