Software - PigeonSuperModel
Pre-trained pigeon models for markerless pose estimation using DeepLabCut and SLEAP.
Welcome to the PigeonSuperModel
This project is hosted in GitLab.
Why a Pigeon Super Model?
Advances in computational neuroethology and markerless pose tracking are making it ever easier for researchers to quantify animal behavior from non-invasive video recordings. Yet, these models still rely on GPUs for heavier computations and model training taking up to several days. With this Pigeon Super Model we provide multiple pre-trained neural networks for out-of-the-box video analysis of pigeon behavior, no previous labeling and training required. A further downside is the missing standards for video recording and analysis, which makes reproducibility across labs somewhat tricky. Instead, with the Pigeon Super Model, we advocate for a standardized set of markers for pigeon tracking and generalizable models across experiments, animals, and camera setups.
Here we make available a dataset of 1151 manually labeled images of different animals in different settings and from different camera angles. We also provide multiple pre-trained models for popular markerless tracking software (i.e., DeepLabCut and SLEAP) to be used out-of-the-box on your own data without any additional configurations. We originally trained these models to generalize well across different experimental setups, using different cameras and different animals, and we found that pre-trained models can be easily re-trained on outlier frames to specialize on any particular data set using pigeons as a model organism.
The Pigeon Super Model was initiated by Guillermo Hidalgo Gadea and Sarah C. Möser. Data for this project was collected by the Biopsychology Team.
This work was carried out at the Institute of Cognitive Neuroscience (IKN), Department of Biopsychology at the Ruhr-University Bochum and supported by the German Research Foundation DFG in the context of funding the Research Training Group “Situated Cognition” (GRK 2185/1). Original: Gefördert durch die Deutsche Forschungsgemeinschaft (DFG) - Projektnummer GRK 2185/1.
Let me know on Twitter if you found this useful or would like to have a more detailed discussion on any of the methods used above.