How to use the PigeonSuperModel

On the next pages we show you how to use the pre-trained models on your own data.

In short, you will have to:

  1. Download your PigeonSuperModel of choice

  2. Analyze your videos (GPU recommended)

  3. Extract and refine outlier frames (optional)

  4. Re-train the model with your expanded dataset (optional)

  5. Submit your new PigeonSuperModel to GitLab (optional)


In the next pages we provide you with commented jupyter notebooks to guide you through the entire process.


We started the PigeonSuperModel in DeepLabCut, so this has been our software of choice for a while now. Only after finishing the first models we decided to make the dataset systems agnostic and include further models, so the entire project may still be a little skewed towards DeepLabCut. Although we certainly recommend DeepLabCut as a great open source toolbox for markerless pose tracking, we want to acknowledge new, very promising approaches being released every few months. Especially, but not exclusively, SLEAP for fast model training and multi-animal tracking, and JARVIS for training 3D-CNN instead of 2D-CNNs with post-hoc triangulation. Let us know what system you prefer and help us include it in the PigeonSuperModel.



This section is coming soon.