OpenPose-Plus: Pose Estimation in the Wild
OpenPose is the state-of-the-art pose estimation algorithm. In its Caffe codebase, data augmentation, training, and neural networks are most hard-coded. They are difficult to be customised. In addition, key performance features such as embedded platform supports and parallel GPU training are missing. All these limitations makes OpenPose, in these days, hard to be deployed in the wild. To resolve this, we develop OpenPose-Plus, a high-performance yet flexible pose estimation framework that offers many powerful features:
- Flexible combination of standard training dataset with your own custom labelled data.
- Customisable data augmentation pipeline without compromising performance
- Deployment on embedded platforms using TensorRT
- Switchable neural networks (e.g., changing VGG to MobileNet for minimal memory consumption)
- High-performance training using multiple GPUs
Note: This project is under active development. Some TODOs are as follows:
- Pose Proposal Networks, ECCV 2018
Custom Model Training
Training the model is implemented using TensorFlow. To run
train.py, you would need to install packages, shown
in requirements.txt, in your virtual environment (Python 3):
pip3 install -r requirements.txt pip3 install pycocotools
train.py automatically download MSCOCO 2017 dataset into
The default model is VGG19 used in the OpenPose paper.
To customize the model, simply changing it in
You can use
train_config.py to configure the training.
config.DATA.train_data can be:
coco: training data is COCO dataset only (default)
custom: training data is your dataset specified by
coco_and_custom: training data is COCO and your dataset
config.MODEL.name can be:
vgg: VGG19 version (default), slow
vggtiny: VGG tiny version, faster
mobilenet: MobileNet version, faster
Train your model by running:
Additional steps for training on Windows
There are a few extra steps to follow with Windows. Please make sure you have the following prerequisites installed:
Download the wget executable and copy it into one of your folders in System path to use the wget command from anywhere. Use the
path command in command line to find the folders. Paste the wget.exe in one of the folders given by
path. An example folder is
pycocotools is not supported by default on Windows. Use the pycocotools build for Windows at here. Instead of
pip install pycocotools, using:
pip install git+https://github.com/philferriere/cocoapi.git#subdirectory=PythonAPI
Visual C++ Build Tools are required by the build. Everything else is the same.
Training using Multiple GPUs
The pose estimation neural network can take days to train.
To speed up the training, we support multiple GPU training while requiring
minimal changes in your code. We use Horovod to support training on GPUs that can spread across multiple machines.
You need to install the OpenMPI in your machine.
We also provide an example script (
scripts/install-mpi.sh) to help you go through the installation.
Once OpenMPI is installed, you can install Horovod python library as follows:
pip3 install horovod
We also have a Python binding for the engine. The current binding relies on the external tf-pose-estimation project. We are working on providing the Python binding for our high-performance C++ implementation. For now, to enable the binding, please build C++ library for post processing by:
See tf-pose for details.
Live Camera Example
You can look at the examples in the
examples folder to see how to use the inference C++ APIs.
./scripts/live-camera.sh will give you a quick review of how it works.
You can use the project code under a free Apache 2.0 license ONLY IF you:
- Cite the TensorLayer paper and this project in your research article if you are an academic user.
- Acknowledge TensorLayer and this project in your project websites/articles if you are a commercial user.