DTI-Clustering

Pytorch implementation of "Deep Transformation-Invariant Clustering" paper (accepted at NeurIPS 2020 as an oral)

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{monnier2020dticlustering,
  title={{Deep Transformation-Invariant Clustering}},
  author={Monnier, Tom and Groueix, Thibault and Aubry, Mathieu},
  booktitle={NeurIPS},
  year={2020},
}

Installation :construction_worker:

1. Create conda environment

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

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

Following script will download affNIST-test and FRGC datasets, as well as our unfiltered Instagram collections associated to #santaphoto and #weddingkiss:

./download_data.sh

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

• affNist source link
• FRGC source link
• IG santaphoto gdrive link
• IG weddingkiss gdrive link

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 prototype evolutions and transformation prediction examples will be saved. Here is an example of learned MNIST prototypes and transformation predictions for a given query image:

Prototypes

Transformation predictions

2. Reproduce our quantitative results on MNIST-test (10 runs)

cuda=gpu_id config=mnist_test.yml tag=dtikmeans ./multi_pipeline.sh

Switch the model name to dtigmm in the config file to reproduce results for DTI GMM. Available configs are:

• affnist_test.yml
• fashion_mnist.yml
• frgc.yml
• mnist.yml
• mnist_1k.yml
• mnist_color.yml
• mnist_test.yml
• svhn.yml
• usps.yml

3. Reproduce our qualitative results on Instagram collections

1. (skip if you already downloaded data using 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.
2. Launch training with cuda=gpu_id config=instagram.yml tag=santaphoto ./pipeline.sh

That's it! You can apply the process to other IG hashtags like #trevifountain or #weddingkiss and discover prototypes similar to:

4. Reproduce our qualitative results on MegaDepth

1. You need to download desired landmarks from the original MegaDepth project webpage, e.g. Florence Cathedral
2. Move images to a datasets/megadepth/firenze/train folder
3. Launch training with cuda=gpu_id config=megadepth.yml tag=firenze ./pipeline.sh

You should end up with 20 learned prototypes and random sample examples in each cluster. To assess the quality of clustering, you can visualized for each cluster, the prototype, random samples and transformed prototypes like:

Further information

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

Follow-ups

• Monnier et al. - Unsupervised Layered Image Decomposition into Object Prototypes (ICCV 2021)

Previous works on deep transformations

• Deprelle et al. - Learning elementary structures for 3D shape generation and matching (NeurIPS 2019)
• Groueix et al. - 3D-CODED: 3D Correspondences by Deep Deformation (ECCV 2018)
• Groueix et al. - AtlasNet: A Papier-Mache Approach to Learning 3D Surface Generation (CVPR 2018)