DeepPoseKit is a software toolkit with a high-level API for 2D pose estimation of user-defined keypoints using deep learning—written in Python and built using Tensorflow and Keras. Use DeepPoseKit if you need:
- tools for annotating images or video frames with user-defined keypoints
- a straightforward but flexible data augmentation pipeline using the imgaug package
- a Keras-based interface for initializing, training, and evaluating pose estimation models
- easy-to-use methods for saving and loading models and making predictions on new data
DeepPoseKit is designed with a focus on usability and extensibility, as being able to go from idea to result with the least possible delay is key to doing good research.
DeepPoseKit is currently limited to individual pose estimation. If individuals can be easily distinguished visually (i.e., they have differently colored bodies or are marked in some way), then multiple individuals can simply be labeled with separate keypoints (head1, tail1, head2, tail2, etc.). Otherwise DeepPoseKit can be extended to multiple individuals by first localizing, tracking, and cropping individuals with additional software such as idtracker.ai, pinpoint, or Tracktor.
Localization (without tracking) can also be achieved with deep learning software like keras-retinanet, the Tensorflow Object Detection API, or MatterPort's Mask R-CNN.
Check out our paper to find out more.
DeepPoseKit is designed for easy use. For example, training and saving a model requires only a few lines of code:
from deepposekit.io import DataGenerator, TrainingGenerator
from deepposekit.models import StackedDenseNet
data_generator = DataGenerator('/path/to/annotation_data.h5')
train_generator = TrainingGenerator(data_generator)
model = StackedDenseNet(train_generator)
model.fit(batch_size=16, n_workers=8)
model.save('/path/to/saved_model.h5')Loading a trained model and running predictions on new data is also straightforward. For example, running predictions on a new video:
from deepposekit.models import load_model
from deepposekit.io import VideoReader
model = load_model('/path/to/saved_model.h5')
reader = VideoReader('/path/to/video.mp4')
predictions = model.predict(reader)- 1. Create an annotation set
- 2. Annotate your data with our built-in GUI (no Colab support)
- 3. Select and train one our pose estimation models including
StackedDenseNet,StackedHourglass,DeepLabCut, andLEAP. - 4. Use the trained model to:
- a) Initialize keypoints for unannotated data for faster annotations with active learning.
- b) Predict on new data and refine the training set to improve performance.
- a) Initialize keypoints for unannotated data for faster annotations with active learning.
- See our example notebooks
- Check the documentation
- Read our paper
DeepPoseKit is designed to be extensible, so loading data in other formats is possible.
If you have annotated data from DeepLabCut (http://deeplabcut.org), try our (experimental) example notebook for loading data in this format.
Have data in another format? You can write your own custom generator to load it.
Check out the example for writing custom data generators.
DeepPoseKit requires Tensorflow for training and using pose estimation models. Tensorflow should be manually installed, along with dependencies such as CUDA and cuDNN, before installing DeepPoseKit:
- Tensorflow Installation Instructions
- Any Tensorflow version >=1.13.0 should be compatible (including 2.0).
DeepPoseKit has only been tested on Ubuntu 18.04, which is the recommended system for using the toolkit.
Install the latest stable release with pip:
pip install --update deepposekitInstall the latest development version with pip:
pip install --update git+https://www.github.com/jgraving/deepposekit.gitYou can download example datasets from our DeepPoseKit Data repository:
git clone https://www.github.com/jgraving/deepposekit-dataTo install DeepPoseKit on Windows, you must first manually install Shapely, one of the dependencies for the imgaug package:
conda install -c conda-forge shapelyWe also recommend installing DeepPoseKit from within Python rather than using the command line, either from within Jupyter or another IDE, to ensure it is installed in the correct working environment:
import sys
!{sys.executable} -m pip install --update deepposekitDeepPoseKit was developed by Jake Graving and Daniel Chae, and is still being actively developed. .
We welcome community involvement and public contributions to the toolkit. If you wish to contribute, please fork the repository to make your modifications and submit a pull request.
If you'd like to get involved with developing DeepPoseKit, get in touch (jgraving@gmail.com) and check out our development roadmap to see future plans for the package.
Please submit bugs or feature requests to the GitHub issue tracker. Please limit reported issues to the DeepPoseKit codebase and provide as much detail as you can with a minimal working example if possible.
If you experience problems with Tensorflow, such as installing CUDA or cuDNN dependencies, then please direct issues to those development teams.
Released under a Apache 2.0 License. See LICENSE for details.
If you use DeepPoseKit for your research please cite our open-access paper:
@article{graving2019deepposekit,
title={DeepPoseKit, a software toolkit for fast and robust animal pose estimation using deep learning},
author={Graving, Jacob M and Chae, Daniel and Naik, Hemal and Li, Liang and Koger, Benjamin and Costelloe, Blair R and Couzin, Iain D},
journal={eLife},
volume={8},
pages={e47994},
year={2019},
publisher={eLife Sciences Publications Limited}
url={https://doi.org/10.7554/eLife.47994},
}
You can also read our open-access preprint.
If you use the imgaug package for data augmentation, please also consider citing it.
If you use data that was annotated with the DeepLabCut package (http://deeplabcut.org) for your research, be sure to cite it.
Please also consider citing the relevant references for the pose estimation model(s) used in your research, which can be found in the documentation (i.e., StackedDenseNet, StackedHourglass, DeepLabCut, LEAP).
- October 2019: Our paper describing DeepPoseKit is published at eLife! (http://paper.deepposekit.org)
- September 2019:
- Nature News covers DeepPoseKit: Deep learning powers a motion-tracking revolution
- v0.3.0 is released. See the release notes.
- April 2019: The DeepPoseKit preprint is on biorxiv (http://preprint.deepposekit.org)