VI-Glow

PyTorch code for our paper: "Color Visual Illusions: A Statistics-based Computational Model", NeurIPS 2020

Preprint version on ArXiv.

Usage

Install Requirements

pip3 install -r requirements.txt

Prepare your data

Download an external dataset, i.e. Places.

This implementation uses torchvision.datasets.ImageFolder. Therefore, the images should be arranged as follows:

+-- <dataset_folder>
|   +-- <class_folder>
|   |   +-- *.png
|   +-- <class_folder>
|   |   +-- *.png

Training

1. Prepare a profile file (such as profile/patch16.json)
2. Run the training script:

python train.py [profile]

You can download a pretrained model from here.

Generate Likelihood Graphs

1. Make a folder with a sequence of input images (structure of torchvision.datasets.ImageFolder)
2. Prepare a profile file (such as profile/patch16_graph.json)
3. Run the script:

python image_patch_graph.py [profile] --output_path <output_folder> --var_name <X label, i.e. Hue>

Generate Likelihood Heatmap

1. Prepare a profile file (such as profile/patch16_heatmap.json)
2. Run the script:

python image_heatmap.py [profile] --output_path <output_folder>

Manipulate Images

1. Generate a mask for your image (it should have the same name as the image but located in another folder).
2. Prepare a profile file (such as profile/patch16_manipulation.json)
3. Run the script:

python image_heatmap.py [profile] --output_path <output_folder> --mask_folder <mask_folder>

Citation

We hope you find our work helpful for your research.

If you are using our code, please consider citing our paper.

@article{hirsch2020color,
  title={Color Visual Illusions: A Statistics-based Computational Model},
  author={Hirsch, Elad and Tal, Ayellet},
  journal={Advances in Neural Information Processing Systems},
  volume={33},
  year={2020}
}

Acknowledgements

This project uses source files of corenel/pytorch-glow.

We also acknowledge the official repository of Glow, by OpenAI.