SATGAN

TODO: update the article URL once published.

Article about this implemention

The self-attentive task generative adversarial network (SATGAN) learns to emulate realistic target sensor noise characteristics in order to augment existing datasets with simulated scenes that better approximate real-world systems. It learns a mapping from random input noise to realistic target-domain sensor characteristics while maintaining semantic information in simulated scenes through the use of a task network. Example real images of a space domain awareness (SDA) scene from the original paper are shown below:

Example noiseless simulated scenes used as context are below:

Finally example simulated scenes with generated addative noise are shown below:

SATGAN comprises three parts: a generator based on a U-net implementation, a discriminator based on PatchGAN, and a task network based on [Fletcher et al.]. The SATGAN architecture is illustrated below:

Setup

Prerequisites

• Tensorflow >= 2.2.1
• Tensorflow-addons >= 0.11.2 (for optional mish activation)
• MISS YOLOv3

Recommended

• Linux with Tensorflow GPU edition + cuDNN

Getting Started

# clone this repo
git clone https://github.com/Engineero/satgan.git
cd satgan

# train the model (this may take 1-8 hours depending on GPU, on CPU you will be waiting for a bit)
python train_satgan.py \
  --mode train \
  --output_dir model_train \
  --max_epochs 200 \
  --input_dir my_data/train \

Citation

TODO: update paper link

If you use this code for your research, please cite the paper this code is based on: Self-attending task generative adversarial network for realistic satellite image creation:

@article{toner_self-attending_2021,
	title = {Self-{Attending} {Task} {Generative} {Adversarial} {Network} for {Realistic} {Satellite} {Image} {Creation}},
	url = {https://arxiv.org/abs/2111.09463v1},
	language = {en},
	urldate = {2021-11-19},
	author = {Toner, Nathan and Fletcher, Justin},
	month = nov,
	year = {2021},
	file = {Snapshot:/Users/nathantoner/Zotero/storage/K7AHTQEU/2111.html:text/html},
}

Acknowledgments