Introduction to DeepPoseKit
A high-level API for 2D pose estimation of user-defined keypoints using deep learning — written in Python and built on TensorFlow / Keras.
DeepPoseKit is an open-source software toolkit for fast and robust 2D animal pose estimation. It is published in eLife (Graving et al., 2019) and is released under the Apache 2.0 License on GitHub.
What it gives you
- Tools for annotating images or video frames with user-defined keypoints.
- A data-augmentation pipeline built on top of
imgaug. - A Keras-style interface to initialise, train and evaluate pose-estimation models.
- Efficient model backbones —
StackedDenseNetandStackedHourglass. - A GPU-based subpixel peak-detection algorithm for extracting keypoint locations.
- Methods for saving / loading models and running predictions on new images or videos.
How it relates to other tools
DeepPoseKit was designed alongside, and benchmarked against, the leading animal pose-estimation toolkits — DeepLabCut and LEAP. On the evaluation datasets in the paper it achieves more than 2× faster inference than these tools with no loss in keypoint accuracy.
The DeepPoseKit family
This documentation site is one of four sibling sites:
- www.deepposekit.org — the project home page.
- docs.deepposekit.org — you are here.
- paper.deepposekit.org — the published eLife article.
- preprint.deepposekit.org — the bioRxiv preprint.
Where to go next
Continue to Installation to set up your environment, then follow the Quickstart to train your first model.