/mcgan-cvprw2017-chainer

This is an implementation of our CVPRW2017 paper "Filmy Cloud Removal on Satellite Imagery with Multispectral Conditional Generative Adversarial Nets".

Primary LanguagePython

Multispectral conditional Generative Adversarial Nets

This repository is an implementation of "Filmy Cloud Removal on Satellite Imagery with Multispectral Conditional Generative Adversarial Nets".

Results

Requirements

I recommend Anaconda to manage your Python libraries.
Because it is easy to install some of the libraries necessary to prepare the data.

  • Python3 (tested with 3.5.4)
  • Chainer (tested with 5.0.0)
  • cupy (tested with 5.0.0)
  • matplotlib (tested with 2.2.2)
  • OpenCV (tested with 3.3.1)
  • tqdm (tested with 4.15.0)
  • PyYAML (tested with 3.12)
  • mpi4py (tested with 3.0.0)

Preparing the data

Please refer to make_dataset/README.md.

Training examples

You need set each parameters in a config file.

CUDA_VISIBLE_DEVICES=0 python train_pix2pix.py --config_path configs/config_nirrgb2rgbcloud.yml --results_dir results/pix2pix

If you want to resume the training from snapshot, use --snapshot option.

  • pretrained model (WIP)

Evaluation examples

CUDA_VISIBLE_DEVICES=0 python test.py --dir_nir <path to nir dir> --dir_rgb <path to rgb dir> --imlist_nir <path to nir list file> --imlist_rgb <path to rgb list file> --results_dir results/test_pix2pix --config_path results/pix2pix/config_nirrgb2rgbcloud.yml --gen_model results/pix2pix/Generator_<iterations>.npz

License

Academic use only.