Shape Generation and Completion Through Point-Voxel Diffusion

Project | Paper

Implementation of Shape Generation and Completion Through Point-Voxel Diffusion

Linqi Zhou, Yilun Du, Jiajun Wu

Requirements:

Install Python environment

module load conda
module load artifactory
mamba env create --file environment.yml
# mamba env update --file environment.yml
conda activate PVD
module load gcc/11.2.0
module load mpfr/4.0.2

Install PyTorchEMD by

cd metrics/PyTorchEMD
python setup.py install
cp build/**/emd_cuda.cpython-*-x86_64-linux-gnu.so .

The code was tested on Unbuntu with Titan RTX.

Data

For generation, we use ShapeNet point cloud, which can be downloaded here. Or at /data/users/lfainsin/ShapeNetCore.v2.PC15k.zip.

For completion, we use ShapeNet rendering provided by GenRe. We provide script convert_cam_params.py to process the provided data.

For training the model on shape completion, we need camera parameters for each view which are not directly available. To obtain these, simply run

$ python convert_cam_params.py --dataroot DATA_DIR --mitsuba_xml_root XML_DIR

which will create ..._cam_params.npz in each provided data folder for each view.

Pretrained models

Pretrained models can be downloaded here. Or at /data/users/lfainsin/PVD/checkpoints.

Training:

$ python train_generation.py --category car|chair|airplane

Please refer to the python file for optimal training parameters.

Testing:

$ python train_generation.py --category car|chair|airplane --model MODEL_PATH

Results

Some generation and completion results are as follows.

Multimodal completion on a ShapeNet chair.

Multimodal completion on PartNet.

Multimodal completion on two Redwood 3DScan chairs.

Reference

@inproceedings{Zhou_2021_ICCV,
  author    = {Zhou, Linqi and Du, Yilun and Wu, Jiajun},
  title     = {3D Shape Generation and Completion Through Point-Voxel Diffusion},
  booktitle = {Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV)},
  month     = {October},
  year      = {2021},
  pages     = {5826-5835}
}

Acknowledgement

For any questions related to codes and experiment setting, please contact Linqi Zhou and Yilun Du.