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

• Chainer (Neural network framework)
  • I'm using master version on GitHub, so
    • git clone https://github.com/pfnet/chainer.git
    • cd chainer; python setup.py install
  • and then, if you want to train networks with GPU,
    • pip install chainer-cuda-deps
• progressbar2
  • pip install progressbar2
  • NOTE: it's not progressbar!

bash scripts/download.sh
python scripts/flic_dataset.py
python scripts/lsp_dataset.py

This script downloads FLIC-full dataset (http://vision.grasp.upenn.edu/cgi-bin/index.php?n=VideoLearning.FLIC) and perform cropping regions of human and save poses as numpy files into FLIC-full directory.

• MPII Human Pose Dataset

• of training images: 18079, # of test images: 6908

  • test images don't have any annotations
  • so we split trining imges into training/test joint set
  • each joint set has

• of training joint set: 17928, # of test joint set: 1991

Just run:

nohup python scripts/train.py > AlexNet_flic_LCN_AdaGrad_lr-0.0005.log 2>&1 &

It is same as:

nohup python scripts/train.py \
--model models/AlexNet_flic.py \
--gpu 0 \
--epoch 1000 \
--batchsize 32 \
--prefix AlexNet_LCN_AdaGrad_lr-0.0005 \
--snapshot 10 \
--datadir data/FLIC-full \
--channel 3 \
--flip 1 \
--size 220 \
--crop_pad_inf 1.5 \
--crop_pad_sup 2.0 \
--shift 5 \
--lcn 1 \
--joint_num 7 \
> AlexNet_LCN_AdaGrad_lr-0.0005.log 2>&1 &

--flip 1 means it performs LR flip augmentation, and --flip 0 does nothing. --lcn 1 means local(should be said "global"?) contrast normalization will be applied.

See the help messages with --help option for details.

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

• Go to result dir of a model
• python ../../scripts/draw_filters.py

python scripts/predict_flic.py \
--model results/AlexNet_2015/AlexNet.py \
--param results/AlexNet_2015/AlexNet_epoch_400.chainermodel \
--datadir data/FLIC-full
--gpu 0 \
--batchsize 128 \
--mode test

python scripts/predict_flic.py \
--model results/AlexNet_2015/AlexNet_flic.py \
--param results/AlexNet_2015/AlexNet_epoch_450.chainermodel \
--mode tile \
--n_imgs 25

cd results/AlexNet_2015
bash ../../scripts/create_anime.sh test_450_tiled_pred.jpg test_450_tiled_label.jpg test_450.gif