CycleConsistentDeformation [Project Page] [Paper] [Talk]

Unsupervised cycle-consistent deformation for shape matching
Thibault Groueix, Matthew Fisher, Vladimir G. Kim , Bryan C. Russell, Mathieu Aubry
In SGP, 2019.

Install

This implementation uses Pytorch and CUDA-10.0.

git clone https://github.com/ThibaultGROUEIX/CycleConsistentDeformation.git
cd CycleConsistentDeformation; git submodule update --init
conda env create -f environment.yml; conda activate sensei
conda install pytorch torchvision cudatoolkit=10.0 -c pytorch
pip install http://imagine.enpc.fr/~langloip/data/pymesh2-0.2.1-cp37-cp37m-linux_x86_64.whl

Demo

1. Target shape + Source Shape

Deform a source in a target and propagate a high-frequency texture from the source to the target, through the reconstruction. -> Require 2GB Gpu memory and 10sec to run (Tesla V100-SXM2-16GB).

chmod +x ./log/download_trained_models.sh # download the trained models
./log/download_pretrained_models.sh # download the trained models
export SHAPENET_PATH=./data/demo/
python inference/forward_source_target.py --cat Chair --shapenetv1_path ${SHAPENET_PATH} --logdir Chair_unsup --eval_source ${SHAPENET_PATH}03001627/7f9733a6370440dcc8687ff9b0b4e4ac/model.obj --eval_target ${SHAPENET_PATH}03001627/257deb231ce652169f2349486c570dd4/model.obj

2. Target shape + Source Dataset

Find the best source candidates for a given target and call forward_source_target() for each. ->Require 3GB Gpu memory and 90sec to run (Tesla V100-SXM2-16GB). First get a download link for shapenetcorev1 by creating an account here -> ShapenetCoreV1

# download Shapenetv1
wget {YOUR_DOWNLOAD_LINK} 
unzip ShapeNetCore.v1.zip; cd ShapeNetCore.v1
unzip 02691156.zip; unzip 03001627.zip; unzip 03636649.zip; unzip 04379243.zip;
export SHAPENET_PATH={YOUR_SHAPENET_PATH}

python inference/get_criterion_shape.py --cat Chair --shapenetv1_path ${SHAPENET_PATH} --logdir Chair_unsup --eval_get_criterions_for_shape ./data/dataset_shapenet/03001627/257deb231ce652169f2349486c570dd4.txt 

Training

python ./training/train_shapenet.py
python ./inference/eval_segmentation.py

Check out the options in auxiliary/argument_parser.py

Reproduce the paper

ERRATUM

The cycle consistency losses of equations 4 and 5 from the paper (2-cycle and 3-cycle consistency) use projection operators which are not differentiable, thus breaking the gradient flow in the cycles. The qualitative/quantitative results hold and we will update the method section of the paper accordingly. Thanks @Jan Bednařík for spotting this!

Citing this work

@INPROCEEDINGS{groueix19cycleconsistentdeformation,
  title     = {Unsupervised cycle-consistent deformation for shape matching},
  author    = {Groueix, Thibault and Fisher, Matthew and Kim, Vova and Russell, Bryan and Aubry, Mathieu},
  booktitle = {Symposium on Geometry Processing (SGP)},
  year      = {2019}
}

License

MIT : This work was funded by Adobe System and Ecole Doctorale MSTIC.