question about v2
XiaoyuShi97 opened this issue · 8 comments
Hi, I find that you mention the model is re-trained. I wonder if the old checkpoint is still compatible with v2 implementation? thanks！
Hi @XiaoyuShi97! Theoretically, the old checkpoint can work with the new codebase, but it is probably going to be worse than the new checkpoint because some components are implemented slightly differently. However, you can still use the old version. Please see https://github.com/facebookresearch/co-tracker?tab=readme-ov-file#previous-version
Thanks for your prompt reply. In my case, another project uses frozen cotracker v1 as a sub-module. But when I substitute it with v2, the performance drops a lot. Is there possible reason for this drop? Or can I re-write v1 following your v2 implementation so that I can continue to use v1 with batch size larger than 1? Thanks!
Hmm, could you tell me what exactly is happening? Maybe you can send a video with predicted tracks if possible?
Yes, it should be relatively easy to use the old checkpoint with the new codebase. You'd need to use the old UpdateFormer without the virtual tracks.
Hi, may I ask if v2 requires more GPU memory or less than v1? @XiaoyuShi97 @nikitakaraevv
Hi @HeadReaper-hc, when tracking more than ~100 points, v2 requires much less GPU memory for the same number of points
@nikitakaraevv get it. Thank you very much.
@nikitakaraevv Hi, another question, does v1 support online mode ?
Hi @1242433760, only if you manage to load the v1 checkpoint in the new codebase.
Hi @XiaoyuShi97! Theoretically, the old checkpoint can work with the new codebase, but it is probably going to be worse than the new checkpoint because some components are implemented slightly differently. However, you can still use the old version. Please see https://github.com/facebookresearch/co-tracker?tab=readme-ov-file#previous-version