Training in low-resolution > inference high-resolution (bad pose estimations!?)

Usage & Issues
#1

Hi everyone,

I’ve trained my network in low-resolution (640x360) videos. When I use it for pose estimations in same-resolution files it does well.
However, when I try to infer poses in 1280x720 videos it gets a mess…
Does anyone have an idea how to resolve this issue?

System:
deeplabcut 2.1.3 (Linux, Ubuntu 16.04)
crop = True
tried with and without dynamic cropping (similar results)

thanks,

Rodrigo

#2

Dear Rodrigo,

that makes sense! Ideally you train the network with similarly sized frames.

If you need broad generalization, you can change the size parameters during augmentation, i.e. scale_jitter_up: 2.5. (see Box 2 Parameters of interest in the network configuration file, pose_cfg.yaml in https://www.nature.com/articles/s41596-019-0176-0)

Cheers,
Alexander

#3

Thanks a lot!
I’ll then try playing with scale_jitter_up and see.

R.

1 Like