Deep-Unsupervised-Pixelization

Paper

Deep Unsupervised Pixelization and Supplementary Material.
Chu Han^, Qiang Wen^, Shengfeng He*, Yinjie Tan, Qianshu Zhu, Guoqiang Han, and Tien-Tsin Wong. (^joint first authors)
ACM Transactions on Graphics (SIGGRAPH Asia 2018 issue), 2018.

Requirement

• Python 3.5
• PIL
• Numpy
• Pytorch 0.4.0
• Ubuntu 16.04 LTS

Dataset

Training Dataset

Create the folders trainA and trainB in the directory ./samples/. Note that trainA and trainB contain the clip arts to be pixelized and pixel arts to be depixelized respectively.

Testing Dataset

Create the folders testA and testB in the directory ./samples/. Note that testA and testB contain the clip arts to be pixelized and pixel arts to be depixelized respectively.

Training

• To train a model:

python3 ./train.py --dataroot ./samples --resize_or_crop crop --gpu_ids 0

or you can directly:

$ bash ./train.sh

You can check the losses of models in the file ./checkpoints_pixelization/loss_log.txt.
More training flags in the files ./options/base_options.py and ./options/train_options.py.

Testing

• After training, all models have been saved in the directory ./checkpoints_pixelization/.
• To test a model:

python3 ./test.py --dataroot ./samples --no_dropout --resize_or_crop crop --gpu_ids 0 --how_many 1 --which_epoch 200

or you can directly:

$ bash ./test.sh

More testing flags in the file ./options/base_options.py.
All testing results will be shown in the directory ./results_pixelization/.

Note

Since this proposed method has been used in commerce, we cannot release the pretrained model and training dataset.

Acknowledgments

Part of the code is based upon pytorch-CycleGAN-and-pix2pix.

Citation

@article{han2018deep,
  title={Deep unsupervised pixelization},
  author={Han, Chu and Wen, Qiang and He, Shengfeng and Zhu, Qianshu and Tan, Yinjie and Han, Guoqiang and Wong, Tien-Tsin},
  journal={ACM Transactions on Graphics (TOG)},
  volume={37},
  number={6},
  pages={1--11},
  year={2018},
  publisher={ACM New York, NY, USA}
}