Custom dataset preparation good practice

[antbertrand]

Hey, I have some questions on how to prepare a custom dataset before using it to train an action recognition network.
I have maybe 1000 hours of video (400x500) at 15 fps, with annotations of 5 different behaviors.

First, is there a recommended fps for action recognition ? I guess a too high or too low fps may make it difficult to identify behaviors.

Secondly, my annotated segments vary in duration from 0,5 sec to maybe 10 seconds. Is it an issue, I found diverging informations about that, some saying that all actions had to last the same duration.

Lastly, would you recommend an annotation format between the video annotation or the ActivityNet annotation format ?

Thank you for the advice,