In "divided_space_time", Why is the class token not used during temporal attention?

[Backdrop9019]

Thank you for your great paper.
I have a question while reading your paper and code.
In the paper, the formula for performing divided_space_time includes the class token during temporal attention, but in the actual code, the cls_token is excluded during temporal attention and included during spatial attention.

This question is a duplicate of the following question, but I did not get a satisfactory answer.
https://github.com/facebookresearch/TimeSformer/issues/74
Is there a reason for this? Would it be possible to obtain experimental results by including the cls_token during temporal attention?

[1998CHC]

Hi! I have the same question.
Have you done any experiments on this problem?