EdgeGAN

Project Page | Paper

SketchyCOCO: Image Generation from Freehand Scene Sketches
Chengying Gao, Qi Liu, Qi Xu, Limin Wang, Jianzhuang Liu, Changqing Zou

Installation

Clone this repo.

git@github.com:sysu-imsl/EdgeGAN.git
cd EdgeGAN

This repo requires TensorFlow 1.14.0 and python 3+.
conda create/activate is suggested to manage multiple versions of tensorflow.
After switching to proper conda environment, run conda install --file requirements.txt

Dataset

Our dataset can be found in SketchyCOCO. Follow the guide and prepare the dataset.

Directory Structure

For singleclass dataset

EdegGAN
└───data
    └───train
    |   |    <file00>.png
    |   |    <file01>.png
    |   |    ...
    |   
    └───test
        |    <file00>.png
        |    <file01>.png
        |    ...

For multiclass dataset

EdegGAN
└───data
    └───train
    |   └───<0>
    |   |    |    <file01>.png
    |   |    |    ...
    |   └───<1>
    |   |   ...
    |   
    └───test
    |   └───<0>
    |   |    |    <file01>.png
    |   |    |    ...
    |   └───<1>
    |   |   ...

For our pretrained model, the class label 0 to 13 correspond to "airplane, cat, giraffe, zebra, dog, elephant, fire hydrant, horse, bicycle, car, traffic light, cow, motorcycle, sheep". Please prepare the input as similar as the examples of training images and test images in images/dataset_example.

Example

Train

Test

Testing

Download the pretrained model from Google Drive trained with 14 classes, and run:

mkdir -p outputs/edgegan
cd outputs/edgegan
cp <checkpoints download path> .
unzip checkpoints.zip
cd ../..

Generate images with models:

python -m edgegan.test --name=edgegan --dataroot=<root of dataset> --dataset=<dataset> --gpu=<gpuid> #(model trained with multi-classes)
python -m edgegan.test --name=[model_name] --dataroot=<root of dataset> --dataset=<dataset> --nomulticlasses --gpu=<gpuid> #(model trained with single class)

the outputs will be located at outputs/edgegan/test_output/ by default

Training

It will cost about fifteen hours to run on a single Nvidia RTX 2080 Ti card.

python -m edgegan.train --name=<new_name> --dataroot=<root of dataset> --dataset=<datsaet_name> --gpu=<gpuid> #(with multi-classes)
python -m edgegan.train --name=<new_name> --dataroot=<root of dataset> --dataset=<datsaet_name> --nomulticlasses --gpu=<gpuid> #(with single class)

Citation

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

@inproceedings{gao2020sketchycoco,
  title={SketchyCOCO: Image Generation From Freehand Scene Sketches},
  author={Gao, Chengying and Liu, Qi and Xu, Qi and Wang, Limin and Liu, Jianzhuang and Zou, Changqing},
  booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
  pages={5174--5183},
  year={2020}
}

sysu-imsl / EdgeGAN