Is there an existing issue for this?

• I have searched the existing issues

Operating System

Windows 11

DeepLabCut version

3.0.0rc9

What engine are you using?

pytorch

DeepLabCut mode

multi animal

Device type

NVIDIA GeForce RTX 4070

Bug description 🐛

I trained dlcrnet-15 and a resnet50 models on >300 epochs with a training dataset of ~1500 frames (95/5) to track 2 identical animals. I enabled the keeping of identities because one of the animals always remains in close viscinity of the unique bodypart (cylinder).

The models seem to accurately track the 2 animal although the mAP and mAR plateau at 70 - 75%, I found that a recurring error appears for all the models I benchmarked . When the mouse outside the cylinder (Expe) gets a bit further from the mouse inside (Stim) or from the cylinder (unique bodypart), the model loses it (see video bellow and poses plot)

I purposefully added frames where the mice are close together to avoid identity switches and frames when they are distant from each other to avoid this issue but it persists.
I wonder if I set a parameter that biases the training somehow, please find attached my config.yaml file.

Steps To Reproduce

config.yaml

Relevant log output

Anything else?

No response

Code of Conduct

• I agree to follow this project's Code of Conduct