NOTE: This is not official implementation. Original paper is DeepPose: Human Pose Estimation via Deep Neural Networks.

• Python 3.5.1+
  • Chainer 1.13.0+
  • numpy 1.9+
  • scikit-image 0.11.3+
  • OpenCV 3.1.0+

I strongly recommend to use Anaconda environment. This repo may be able to be used in Python 2.7 environment, but I haven't tested.

pip install chainer
pip install numpy
pip install scikit-image
# for python3
conda install -c https://conda.binstar.org/menpo opencv3
# for python2
conda install opencv

bash datasets/download.sh
python datasets/flic_dataset.py
python datasets/lsp_dataset.py
python datasets/mpii_dataset.py

• FLIC-full dataset
• LSP Extended dataset
• MPII dataset
  • Annotation
  • Images

• MPII Human Pose Dataset
• training images: 18079, test images: 6908
  • test images don't have any annotations
  • so we split trining imges into training/test joint set
  • each joint set has
• training joint set: 17928, test joint set: 1991

Starting with the prepared shells is the easiest way. If you want to run train.py with your own settings, please check the options first by python scripts/train.py --help and modify one of the following shells to customize training settings.

bash shells/train_flic.sh

bash shells/train_lsp.sh

bash shells/train_mpii.sh

• AlexNet
  • batchsize: 128 -> about 2870 MiB
  • batchsize: 64 -> about 1890 MiB
  • batchsize: 32 (default) -> 1374 MiB
• ResNet50
  • batchsize: 32 -> 6877 MiB

Will add some tools soon