GeoGAN

This is PyTorch implementation of a GeoGAN network for for dense labeling of RGB+IR optical imagery. The code is based on the pytorch-CycleGAN-and-pix2pix

GeoGAN: Project | Paper

Prerequisites

• Linux or macOS
• Python 2 or 3
• CPU or NVIDIA GPU + CUDA CuDNN

Getting Started

Installation

• Install PyTorch 0.4, torchvision, and other dependencies from http://pytorch.org
• Install python libraries visdom and dominate.

pip install visdom dominate

• Alternatively, all dependencies can be installed by

pip install -r requirements.txt

• Clone this repo:

git clone https://github.com/vlkniaz/GeoGAN.git
cd GeoGAN

• For Conda users, we include a script ./scripts/conda_deps.sh to install PyTorch and other libraries.

GeoGAN train/test

• Download a GeoGAN dataset (e.g. maps):

bash ./datasets/download_geogan_dataset.sh maps

• Train a model:

#!./scripts/train_geogan.sh
python train.py --dataroot ./datasets/maps --name maps_geogan --model geo_gan

• To view training results and loss plots, run python -m visdom.server and click the URL http://localhost:8097. To see more intermediate results, check out ./checkpoints/maps_geogan/web/index.html
• Test the model:

#!./scripts/test_geogan.sh
python test.py --dataroot ./datasets/maps --name maps_geogan --model geo_gan

The test results will be saved to a html file here: ./results/maps_geogan/latest_test/index.html.

Apply a pre-trained model (GeoGAN)

• You can download a pretrained model (e.g. isprs) with the following script:

bash ./scripts/download_geogan_model.sh isprs

The pretrained model is saved at ./checkpoints/{name}_pretrained/latest_net_G.pth. The available model is isprs.

• To test the model, you also need to download the isprs dataset:

bash ./datasets/download_geogan_dataset.sh horse2zebra

• Then generate the results using

python test.py --dataroot datasets/horse2zebra/testA --name horse2zebra_pretrained --model test

The option --model test is used for generating results of GeoGAN only for one side. python test.py --model geo_gan will require loading and generating results in both directions, which is sometimes unnecessary. The results will be saved at ./results/. Use --results_dir {directory_path_to_save_result} to specify the results directory.

Datasets

Download GeoGAN datasets and create your own datasets.

Training/Test Tips

Best practice for training and testing your models.

Citation

If you use this code for your research, please cite our papers.

@inproceedings{Kniaz:2018fq,
author = {Kniaz, V V},
title = {{Conditional GANs for Semantic Segmentation of Multispectral Satellite Images}},
booktitle = {SPIE Remote Sensing},
year = {2018},
publisher = {SPIE},
month = sep
}

Related Projects

CycleGAN-Torch | pix2pix-Torch | pix2pixHD | iGAN | BicycleGAN

Acknowledgments

Code is inspired by pytorch-DCGAN.