Question about the number of timesteps
XChenZ opened this issue · comments
Thanks for your impressive job!
I have a question about the impact of the number of timesteps. In the paper, you mention that only three frames(t-20, t, t+20) are used for the temporal transformer. Why not utilize more frames to integrate temporal information? Have you ever experimented with such ablation studies? I'd appreciate it if you share more insight about this.
BTW, looking forward to the code release!
I am really sorry for the late reply!
The reason why I did not integrate more time frames was due to the memory limitation. So it can be one interesting research direction - how to select the frames that can be most helpful to you among the many frames!
The code will be released in a few days :)
Hi! Thanks for your amazing work!
I also have some questions about the time_steps
. I find that time_steps
is set to 1
in all config files, and I cannot find the code for temporal transformer. Is it implemented in current version?