How do you compare with others as your work is trained in an end-to-end way

[Tristan-J]

Hi,
Thanks for your solid theoretical and experiment proof. But as mentioned in the paper, your work is trained in an end-to-end way while others, e.g. mlad, used pretrained i3d which is only on kinetics400. Have you tried with the pretrained i3d as well as rgb-optical_flow two stream, and how about the results? Thanks!