Code accompanying the GCPR 2020 paper

Unsupervised Part Discovery by Unsupervised Disentanglement
Sandro Braun, Patrick Esser, Björn Ommer

arXiv | BibTeX | Project Page

• Requirements
• Training
  • Data
• Evaluation
• Pretrained Models
• BibTeX

A suitable conda environment named braun20parts can be created and activated with:

conda env create -f environment.yaml
conda activate braun20parts

Clone the repo with all it's submodules

git clone --recursive -j8 [email protected]:CompVis/unsupervised-part-segmentation.git

1. For running experiments into the respective subfolders deepfashion, cub and pennaction.
2. Experiments can be run using edflow.

edflow -t xxx/<config.yaml>

• The CUB dataset is the same as in Lorenz19, but we manually added semantic part segmentations and added them in the repo.

1. baseline models with pretrained checkpoints on all datasets can be found in folder baselines
2. evaluation scripts and notebooks can be found in folder evaluation

pretrained models can be found in the respective folder, under train/checkpoints

@inproceedings{braun2020parts,
  title={Unsupervised Part Discovery by Unsupervised Disentanglement},
  author={Braun, Sandro and Esser, Patrick and Ommer, Bj{\"o}rn},
  booktitle={Proceedings of the German Conference on Computer Vision},
  year={2020}
}