facebookresearch/co-tracker

Full video in memory

bhack opened this issue · 2 comments

bhack commented

Is there a way to avoid to load the full video in memory.
Specially cause currently the full video is loaded in memory at full-res before you rescale it internally with
video = F.interpolate(video, tuple(self.interp_shape), mode="bilinear")

nikitakaraevv commented

Hi @bhack, you can resize the video to the interpolation resolution (384x512) before feeding it to the model. If you wish to obtain trajectories in the original resolution, simply multiply them by the ratio between the original and resized resolutions.

bhack commented

If you wish to obtain trajectories in the original resolution, simply multiply them by the ratio between the original and resized resolutions.

This was already done by the current code in the repository.
I was looking just for an ablation at higher resolution.