Jiayi Liu, Hou In Ivan Tam, Ali Mahdavi-Amiri, Manolis Savva
CVPR 2024
Page | Paper | Data (alternative link for data: OneDrive)
We recommend the use of miniconda to manage system dependencies. The environment was tested on Ubuntu 20.04.4 LTS.
# Create a conda environment
conda create -n cage python=3.10
conda activate cage
# Install pytorch
conda install pytorch torchvision torchaudio pytorch-cuda=11.8 -c pytorch -c nvidia
# Install PyGraphviz
conda install --channel conda-forge pygraphviz
# Install other packages
pip install -r requirements.txt
# Install PyTorch3D (not required for training):
conda install -c fvcore -c iopath -c conda-forge fvcore iopath
conda install pytorch3d -c pytorch3d
We share the training data (here~101MB) preprocessed from PartNet-Mobility dataset. Once downloaded, extract the data
and put it directly in the project folder. The data root can be configured with system.datamodule.root=<path/to/your/data/directory>
in configs/cage.yaml
file. If you find it slow to download the data from our server, please try this alternative link on OneDrive.
The preprocessing script can be found in preprocess.py
. We use this script to process the selective categories of objects used in our experiments.
We share the pretrained model (here~80MB) so you can try our demo real quick. Once downloaded, extract the zip file and put it under <project folder>/exps
folder. Since our part retrieval relies on the meshes in the dataset, the data should be already downloaded and put under the project folder by default. Run python demo.py
to start the demo (with a single GPU is preferred). Please see demo.py
for further instructions on the script arguments.
Run python main.py --config configs/cage.yaml --log_dir <folder/for/logs>
to train the model from the scratch. The experiment files will be recorded at ./<log_dir>/cage/<version>
. The original model was trained on two NVIDIA A40 GPUs.
Please cite our work if you find it helpful:
@inproceedings{liu2024cage,
title={CAGE: Controllable Articulation GEneration},
author={Liu, Jiayi and Tam, Hou In Ivan and Mahdavi-Amiri, Ali and Savva, Manolis},
booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
pages={17880--17889},
year={2024}
}
This implementation is partially powered by 🤗Diffusers.