Patch Embedding Weights

Love the work!

I noticed that the patch embedding weights for the vggt backbone in the provided cowtracker checkpoint are different from the patch embedding weights I get when I use VGGT.from_pretrained("facebook/VGGT-1B"). I see the paper says you guys froze the patch embedding layers and only fine tuned other parts. Did you guys use a different VGGT checkpoint?

Thanks!

Start by comparing the patch embedding weights in the provided cowtracker checkpoint with those produced by VGGT.from_pretrained("facebook/VGGT-1B"). Review the paper's statement that patch embedding layers were frozen and determine whether a different VGGT checkpoint was used; done means documenting the checkpoint provenance and explaining the weight difference.

Written by the indexing model from the issue text.