with transfer learning - super animal top view cannot create labeled video

[mrc-sala]

Is there an existing issue for this?

• I have searched the existing issues

Bug description

Hi everyone,

I'm trying to use the SuperAnimal-TopViewMouse to track a multi-animal video. I retrained the model as we use different body parts with "Transfer Learning." Indeed, when I analyze a sample video, I can get a CSV with tracked body parts for each mouse. However, when I create the video, I only get a video with red boxes around the mice, despite selecting body parts in the GUI. Is this a bug, or how can I get a video with the dots for each body part?

Thanks,
M

Operating System

windows 11 using conda and pytorch

DeepLabCut version

dlc version: DLC 3.0.0rc6

DeepLabCut mode

multi animal

Device type

gpu NVIDIA RTX-A2000

Steps To Reproduce

No response

Relevant log output

when try it on ipython reepl:

In [8]: deeplabcut.create_labeled_video(config_path, video_path, draw_skeleton=True, save_frames=False, displayedindividuals="bodyparts")
Starting to process video: C:\Users\sala\Projects\DLC\mouse_interaction\videos\video0025 23-03-27 17-32-53.mp4
Loading C:\Users\sala\Projects\DLC\mouse_interaction-sala-2025-02-28\videos\video0025 23-03-27 17-32-53.mp4 and data.
Duration of video [s]: 335.37, recorded with 30.0 fps!
Overall # of frames: 10061 with cropped frame dimensions: 640 480
Generating frames and creating video.
C:\Users\sala\.conda\envs\DEEPLABCUT\lib\site-packages\deeplabcut\utils\make_labeled_video.py:146: FutureWarning: DataFrame.groupby with axis=1 is deprecated. Do `frame.T.groupby(...)` without axis instead.
  Dataframe.groupby(level="individuals", axis=1).size().values // 3

Anything else?

No response

Code of Conduct

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

[maximpavliv]

Hello @mrc-sala ,

What likelihood values do you have in your .csv analysis file and what pcutoff do you set when creating the labeled video?
Could you try again but set the pcutoff to 0, to check if this is the reason?

[mrc-sala]

Hello @maximpavliv,

Thanks! That was indeed the problem. It seems the training didn’t work correctly. Now, I see one (sometimes two) large points in the video. I guess the various body parts are superimposed—maybe I’ll retry the training.

[maximpavliv]

Yes, looks like the training didn't do well, regarding this particular frame at least.
Have you trained the detector [sufficiently] as well? On this frame we can see that the detector model is already giving incorrect bboxes, so the pose model can only fail (as it will look for one animal in each bounding box).

I will close this issue for now, as the Github Issues page is rather for bugs reports and feature requests. If you struggle to get better performances with the training - please create a topic on DeepLabCut's forum, and the DeepLabcut community or a DeepLabcut developer will give you some tips there 😃