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 customized. 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.
• Customizable 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

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 dataset/coco17. The default model is VGG19 used in the OpenPose paper. To customize the model, simply changing it in models.py.

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 config.DATA.your_xxx
• 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:

python3 train.py

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:

• git
• Visual C++ Build Tools
• wget

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 C:\Windows.

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.

Distributed Training

The pose estimation neural network can take days to train. To speed up training, we support distributed GPU training. We use the KungFu library to scale out training. KungFu is very easy to install and run (compared to the previously used Horovod library which depends on OpenMPI), and simply follow the instruction.

In the following, we assume that you have added kungfu-run into the $PATH.

(i) To run on a machine with 4 GPUs:

kungfu-run -np 4 python3 train.py --parallel --kf-optimizer=sma

The default KungFu optimizer is sma which implements synchronous model averaging. You can also use other KungFu optimizers: sync-sgd (which is the same as the DistributedOptimizer in Horovod) and async-sgd if you train your model in a cluster that has limited bandwidth and straggelers.

(ii) To run on 2 machines (which have the nic eth0 with IPs as 192.168.0.1 and 192.168.0.2):

kungfu-run -np 8 -H 192.168.0.1:4,192.168.0.1:4 -nic eth0 python3 train.py --parallel --kf-optimizer=sma

High-performance Inference using TensorRT

Real-time inference on resource-constrained embedded platforms is always challenging. To resolve this, we provide a TensorRT-compatible inference engine. The engine has two C++ APIs, both defined in include/openpose-plus.hpp. They are for running the TensorFlow model with TensorRT and post-processing respectively.

You can build the APIs into a standard C++ library by just running make pack, provided that you have the following dependencies installed

• tensorRT
• opencv
• gflags

We are improving the performance of the engine. Initial benchmark results for running the full OpenPose model are as follows. On Jetson TX2, the inference speed is 13 frames / second (the mobilenet variant is even faster). On Jetson TX1, the speed is 10 frames / second. On Titan 1050, the speed is 38 frames / second.

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:

./scripts/install-pafprocess.sh

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. Running ./scripts/live-camera.sh will give you a quick review of how it works.

License

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.

Related Discussion

• TensorLayer Slack
• TensorLayer WeChat
• TensorLayer Issues 434
• TensorLayer Issues 416