Models
Contents
Models¶
Here, we describe the pre-trained models we provide based on our PigeonSuperModel dataset. We trained multiple DeepLabCut models based on different architectures (resnet-50, resnet-101, resnet-152) and compared tracking performance across different training stages. We make the final, pre-trained models available for out-of-the-box analysis of new videos. Moreover, we used the same dataset to train UNet models in SLEAP and benchmark performance differences, training and inference rates.
Out-of-the-box Model Performance¶
Video of new subject in known experimental setting¶
New video of known subject in known experimental setting¶
Video of new subject in new experimental setting¶
DeeplabCut¶
Using DeepLabCut, we trained three different network architectures to compare training and inference times, as well as overall tracking performance on known and unseen frames.
ResNet-50¶
Coming soon! We are preparing a short pre-print describing model training and configuration parameters. In the meantime, use this link to download the pre-trained model.
SLEAP¶
Note
This section is coming soon, we would like to polish the pre-print first.
ResNet¶
Coming soon! We are preparing a short pre-print describing model training and configuration parameters.
UNet¶
Coming soon! We are preparing a short pre-print describing model training and configuration parameters.