DeepPose

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

Requirements

• Chainer 1.5+ (Neural network framework)
• numpy 1.9+
• scipy 0.16+
• scikit-learn 0.15+
• OpenCV 2.4+

Data preparation

bash shells/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 Dataset

• 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

Start training

For FLIC Dataset

Just run:

nohup python scripts/train.py > AlexNet_flic.log 2>&1 < /dev/null &

For speed:

CHAINER_TYPE_CHECK=0 nohup python scripts/train.py > AlexNet_flic.log 2>&1 < /dev/null &

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.

GPU memory requirement

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

Visualize Filters of 1st conv layer

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

Visualize Prediction

Example

Prediction and visualize them and calc mean errors

python scripts/evaluate_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

Tile some randomly selected result images

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

Create animated GIF to intuitively compare predictions and labels

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