Winding Clearness for Differentiable Point Cloud Optimization

This document contains the source code for our work “Winding Clearness for Differentiable Point Cloud Optimization”.

• The ./optimization folder includes the code for single point cloud optimization proposed in Section 5.1.
• The ./learning folder includes the code for incorporating winding clearness as a geometric constraint in the standard DDPM proposed in Section 5.2.

The code execution requires the following dependencies. These versions are not specified. Other compatible versions may also work, provided that the torch versions align with the available cuda toolkit.

python==3.10
pytorch==2.0.0
torchvision==0.15.1
cudatoolkit==11.8
tqdm==4.32.1
open3d==0.16.0
trimesh=4.0.8
scipy==1.11.4

These dependencies are sufficient for the single point cloud optimization approach. For the diffusion-based approach, PyTorchEMD is required and can be installed with the following commands:

cd metrics/PyTorchEMD
python setup.py install
cp build/**/emd_cuda.cpython-310m-x86_64-linux-gnu.so .

Please ensure that the gcc version aligns with the system configuration as required in "learning/metrics/ChamferDistancePytorch/chamfer3D/setup.py" and "learning/modules/functional/backend.py"

Execute the following commands to run the single point cloud optimization method:

python optimization_based_2D.py --dataroot ./input/2D --result_label 2D
python optimization_based_3D.py --dataroot ./input/3D --result_label 3D 
python optimization_based_3D.py --dataroot ./input/3D2 --result_label 3D2 --lambda_k 20.0
python optimization_based_3D_multi_batch.py --dataroot ./input/3D_multi_batch --result_label 3D_multi_batch

The last row of the command can run the multi-batch method stated in Section 7, where the point cloud contains 50K points. Results will be stored in the "output" folder.

The baseline of the diffusion-based method is PVD. All relevant data and pretrained models are accessible or can be trained as described in the GitHub page of PVD. We include here the fine-tuned models of our method for the chair category and all the 55 classes at "./new_joint_train_model_chair" and "./new_joint_train_model_all", respectively.

The generated point clouds for "PVD" and "PVD+Ours" of the chair category is in "./results".

To run the train and test code, the ShapeNet point cloud are requires, which can be downloaded here. After putting pretrained model "chair_1799.pth" of PVD into "./ckpt/generation", we can run the following command to generate the results of "PVD":

python test_generation_chair.py

and the following command to generate the results of "PVD+Ours":

python test_generation_regularization_chair.py

The generated results will be located in "output/test_generation_regularization_chair". The joint training strategy of our method can be executed by:

python train_generation_joint_chair.py

@article{XIAO2025103930,
title = {Winding clearness for differentiable point cloud optimization},
journal = {Computer-Aided Design},
volume = {188},
pages = {103930},
year = {2025},
issn = {0010-4485},
doi = {https://doi.org/10.1016/j.cad.2025.103930},
url = {https://www.sciencedirect.com/science/article/pii/S0010448525000910},
author = {Dong Xiao and Yueji Ma and Zuoqiang Shi and Shiqing Xin and Wenping Wang and Bailin Deng and Bin Wang},
}