GPU out of memory about co-tracker

[nikitakaraevv]

Hi @fenaux, it looks like you were allocated a GPU in Colab that doesn't have enough memory for this part of the demo. You can try reducing grid_size when estimating tracks with a segmentation mask. I'll fix this later in the notebook.

[fenaux]

Thanks for your fast answer,
on my computer (GPU is 16 GiB as the standard one on Colab) it works with grid_size = 60 (I did not try more)
once this change made no more problem for the dense tracks

Relationship between grid_size and pred_tracks.size() is not clear to me
with grid_size = 40 pred_tracks.size() = [1, number of frames, 161, 2]
and with grid_size = 60 pred_tracks.size() = [1, number of frames, 381,2]

[nikitakaraevv]

In the example with segmentation, we sample grid_size * grid_size points on a regular grid. Later, these points are filtered using a binary segmentation mask, preserving only the points that lie inside the segmented area. This leaves us with 161 points out of 40 * 40 = 1600 initially sampled points and 381 out of 60 * 60 = 3600 initially sampled points.

[fenaux]

Once again thanks for your prompt and clear answer