Awesome

DTI-Sprites

Pytorch implementation of "Unsupervised Layered Image Decomposition into Object Prototypes" paper (accepted to ICCV 2021)

Check out our paper and webpage for details!

If you find this code useful, don't forget to star the repo ⭐ and cite the paper:

@inproceedings{monnier2021dtisprites,
  title={{Unsupervised Layered Image Decomposition into Object Prototypes}},
  author={Monnier, Tom and Vincent, Elliot and Ponce, Jean and Aubry, Mathieu},
  booktitle={ICCV},
  year={2021},
}

Installation :construction_worker:

1. Create conda environment

conda env create -f environment.yml
conda activate dti-sprites

Optional: some monitoring routines are implemented, you can use them by specifying the visdom port in the config file. You will need to install visdom from source beforehand

git clone https://github.com/facebookresearch/visdom
cd visdom && pip install -e .

2. Download non-torchvision datasets

./download_data.sh

This command will download following datasets:

Tetrominoes, Multi-dSprites and CLEVR6 (link to the original repo multi-object datasets with raw tfrecords)
GTSRB (link to the original dataset page)
Weizmann Horse database (link to the original dataset page)
Instagram collections associated to #santaphoto and #weddingkiss (link to the original repo with datasets links and descriptions)

NB: it may happen that gdown hangs, if so you can download them by hand with following gdrive links, unzip and move them to the datasets folder:

How to use :rocket:

1. Launch a training

cuda=gpu_id config=filename.yml tag=run_tag ./pipeline.sh

where:

gpu_id is a target cuda device id,
filename.yml is a YAML config located in configs folder,
run_tag is a tag for the experiment.

Results are saved at runs/${DATASET}/${DATE}_${run_tag} where DATASET is the dataset name specified in filename.yml and DATE is the current date in mmdd format. Some training visual results like sprites evolution and reconstruction examples will be saved. Here is an example from Tetrominoes dataset:

Reconstruction examples

Sprites evolution and final

More visual results are available at https://imagine.enpc.fr/~monniert/DTI-Sprites/extra_results/.

2. Reproduce our quantitative results

To launch 5 runs on Tetrominoes benchmark and reproduce our results:

cuda=gpu_id config=tetro.yml tag=default ./multi_pipeline.sh

Available configs are:

Multi-object benchmarks: tetro.yml, dsprites.yml, clevr6.yml
Clustering benchmarks: gtsrb8.yml, svhn.yml
Cosegmentation dataset: horse.yml

3. Reproduce our qualitative results on Instagram collections

(skip if already downloaded with script above) Create a santaphoto dataset by running process_insta_santa.sh script. It can take a while to scrape the 10k posts from Instagram.
Launch training with cuda=gpu_id config=instagram.yml tag=santaphoto ./pipeline.sh

That's it! Overall you should end up with the following qualitative results:

Automatically discovered sprites

Further information

If you like this project, please check out related works on deep transformations from our group: