Frame Numbers in Trained Model?

Hello,

Is it possible to learn frame numbers in the released pre-trained models? In the released model, What I understand is that there are 2^4 temporal stride which makes 16 frame numbers. However, we know that if the model is trained with more frame numbers, the performance can increase significantly which can be observed from the paper "Can Spatiotemporal 3D CNNs Retrace the History of 2D CNNs and ImageNet?" and you released the paper with the results of 64 frames. Therefore, if the released pre-trained models are trained with 16 frames, could you please release the pre-trained versions of 64 frame?

One more question is that if you trained the released architecture with 64 frame, what changes did you make? Did you just simply take the mean of the temporal dimension?

Thanks in advance