Triggering torch.isnan(coords).any()

[Jafonso-sudo]

Hello,

On seemingly larger inputs, I seem to trigger the condition in line 520 of spatracker.py (in the forward_iteration function). I imagine you have encountered it too for the condition to be there. Do you have any idea what triggers it and how to fix it?

Thanks in advance!

[zhangzjjjjjj]

@Jafonso-sudo @henry123-boy Hello, I've encountered the similar issue, which seems to happen randomly. Have you come up with any ideas?

[qinghew]

@Jafonso-sudo @zhangzjjjjjj Hello, I've encountered the similar issue, which seems to happen randomly. Have you come up with any ideas?

[Jafonso-sudo]

Unfortunately I gave up and decided to use CoTracker 3 when it was released. It's really unfortunate since the problem I'm applying it to is inherently 3D, so I am sure SpatialTracker could have been beneficial :(

[henry123-boy]

I sincerely apologize for being extremely busy over the past few months. Could someone kindly provide a video sample where the NAN issue occurs?

[Jafonso-sudo]

Hey @henry123-boy, no need to apologize at all, you owe as nothing. I'll try to see if I have time to try and setup SpatialTracker again in the next few days and get you a sample. Though, I imagine someone on your team encountered the same problem since the breakpoint condition is already pushed to the repository.