PoNA

Code for Our TIP paper.

Data preparation

#environment setup
conda create -n tip python=3.6
conda activate tip
conda install pytorch=1.2 cudatoolkit=10.0 torchvision
pip install scikit-image pillow pandas tqdm dominate 

The process of data preparation is following Pose Transfer

Testing

Here is the results（baidu fetch code:abcd) if you are interested in this work.

Checkpoint of DeepFashion is here (fetch code: abcd). You can test by yourself.

Note that the Checkpoint of Market1501 is lost due to COVID-19 and the architecture is a little difference compared with DeepFashion. We will release it as soon as possible.

#fashion
python test.py --dataroot ./fashion_data/ --name fashion_tip --model PoNA --phase test --dataset_mode keypoint --norm instance --batchSize 1 --resize_or_crop no --gpu_ids 0 --BP_input_nc 18 --no_flip --which_model_netG PoNA --checkpoints_dir ./checkpoints/ --pairLst ./fashion_data/fasion-resize-pairs-test.csv --which_epoch test1 --results_dir ./tip_fashion --display_id 0

Evaluation

#environment setup
conda create -n test_tip python=3.6
conda activate test_tip
pip install -r requirements_tf.txt
# test script
python tool/getMatric_fashion.py

Citation

If you use this code, please cite our paper.

@ARTICLE{9222550,
  author={K. {Li} and J. {Zhang} and Y. {Liu} and Y. -K. {Lai} and Q. {Dai}},
  journal={IEEE Transactions on Image Processing}, 
  title={PoNA: Pose-Guided Non-Local Attention for Human Pose Transfer}, 
  year={2020},
  volume={29},
  number={},
  pages={9584-9599},
  doi={10.1109/TIP.2020.3029455}}

Acknowledgments

Our code is based on Pose Transfer.