facebookresearch / co-tracker

CoTracker is a model for tracking any point (pixel) on a video.
https://co-tracker.github.io/
Other
2.71k stars 194 forks source link

question about windowed inference #74

Closed XiaoyuShi97 closed 6 months ago

XiaoyuShi97 commented 6 months ago

image

Hi, thanks for sharing this great work. About the windowed inference, I understand that the first T/2 components can be initialized from the last T/2 components from the previous window as overlap exists. But what about the last T/2? Specifically, "copies of the last time t=T/2-1" from P^{(M,1)}", what does it mean?

XiaoyuShi97 commented 6 months ago

Do you refer to "replicate" mode of padding?

nikitakaraevv commented 6 months ago

Hi @XiaoyuShi97, yes, that exactly means the "replicate" padding mode: https://github.com/facebookresearch/co-tracker/blob/0f9d32869ac51f3bd12c5ead9c206366cfb6caea/cotracker/models/core/cotracker/cotracker.py#L295

XiaoyuShi97 commented 6 months ago

Now I fully understand this point. It makes sense. Thanks!