BUG: keypoints predictions often outside image

[phanhuynh]

Version information:

facemap version: 1.0.6 platform: win32 python version: 3.8.18 torch version: 2.1.0+cpu numpy version: 1.24.4

Describe the issue:

When I process keypoints, even without finetuning, I find that facemap sometimes places the keypoints outside of the image. So when I play the recording, the whole frame shifts in size in order to capture all of the keypoints.

Also (since I'm not sure where pose this question): does it matter if I use jpeg or png compression prior to importing the recording?

Error message:

No response