ST-GAN: Spatial Transformer Generative Adversarial Networks for Image Compositing

Chen-Hsuan Lin, Ersin Yumer, Oliver Wang, Eli Shechtman, and Simon Lucey
IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2018

Project page: https://chenhsuanlin.bitbucket.io/spatial-transformer-GAN
Paper: https://chenhsuanlin.bitbucket.io/spatial-transformer-GAN/paper.pdf
arXiv preprint: https://arxiv.org/abs/1803.01837

We provide TensorFlow code for the following experiments:

• indoor object compositing
• glasses compositing

---

Prerequisites

This code is developed with Python3 (python3). TensorFlow r1.0+ is required. The dependencies can install by running

pip3 install --upgrade numpy scipy imageio termcolor tensorflow-gpu

If you don't have sudo access, add the --user flag.

Dataset (indoor objects)

The dataset (6.4GB) can be downloaded by running the command

wget https://cmu.box.com/shared/static/us4vjubhcgt5ziiikaw8kyowut1sopy5.gz

After downloading, run tar -zxf us4vjubhcgt5ziiikaw8kyowut1sopy5.gz under indoor. The files will be extracted to a directory dataset.

Dataset (glasses)

The following from the CelebA dataset is required:

• Aligned & cropped images
• Attribute annotations
• Train/val/test partitions

After downloading CelebA, run python3 preprocess_celebA.py under glasses to convert the train/test split to .npy format.
(The variable celebA_path should be changed to the path where CelebA is downloaded)

Pretrained model (indoor objects)

The pretrained model (89.6MB) can be downloaded by running the command

wget https://cmu.box.com/shared/static/ygl08wfsc2omutwrvra3u7zjjsu1ovwv.gz

After downloading, run tar -zxf ygl08wfsc2omutwrvra3u7zjjsu1ovwv.gz under indoor. The files will be extracted to a directory models_0.

Pretrained model (glasses)

The pretrained model (121MB) can be downloaded by running the command

wget https://cmu.box.com/shared/static/5ad2lbjuvze9iey2up6hcisg6dctii4h.gz

After downloading, run tar -zxf 5ad2lbjuvze9iey2up6hcisg6dctii4h.gz under glasses. The files will be extracted to a directory models_0.

Running the code

To train ST-GAN, run ./train.sh under indoor/glasses.
The checkpoints are saved in the automatically created directory model_GROUP; summaries are saved in summary_GROUP.
The list of optional arguments can be found by executing python3 train_STGAN.py --help.

To evaluate ST-GAN, run ./test.sh under indoor/glasses.
The output of ST-GAN will be saved in the directory eval_GROUP (automatically created).
For glasses, you can also replace the --loadImage flag with the path to your file if you want to try with your own images (the image should be of size 144x144x3).

Visualizing the results

We've included code to visualize the training over TensorBoard. To execute, run

tensorboard --logdir=summary_GROUP --port=6006

We provide three types of data visualization:

1. SCALARS: training curves over iterations (not much meaningful)
2. IMAGES: composite results
3. GRAPH: network architecture

---

If you find our code useful for your research, please cite

@inproceedings{lin2018stgan,
  title={ST-GAN: Spatial Transformer Generative Adversarial Networks for Image Compositing},
  author={Lin, Chen-Hsuan and Yumer, Ersin and Wang, Oliver and Shechtman, Eli and Lucey, Simon},
  booktitle={IEEE Conference on Computer Vision and Pattern Recognition ({CVPR})},
  year={2018}
}

Please contact me (chlin@cmu.edu) if you have any questions!