NerVE

Official PyTorch implementation for paper NerVE: Neural Volumetric Edges for Parametric Curve Extraction from Point Cloud.

Paper

Requirements

Install from requirements.txt.

pip install -r requirements.txt

Build the avg_pooling module. NOTE: the CUDA version of your installed PyTorch has to be consistent with the version of your nvcc.

cd network/avg_pooling
python setup.py build_ext
python setup.py install

Blender, for visualization.

Quick Start

For a quick and easy try, we provide the data for demo and pre-trained model checkpoints:

demo_data_checkpoint.zip

Unzip the downloaded file at the root, then get demo/(demo data) and results/(checkpoints). In the folder demo/, there are point cloud files of 4 CAD models from ABC dataset. Run the following code to obtain the NerVE prediction of network and corresponding piecewise-linear(PWL) curves:

python demo_predict_pwl.py

If everything goes well, we now have a file named pred_nerve_pwl_curve.pkl in the folder of each CAD model. Next convert these PWL curves to parametric curves by running:

python utils/pwl2CAD/CAD_curve.py

Finally, render the parametric curves and NerVE PWL curves with Blender:

blender -b --python utils/visualization_blender/batch_render_curve.py

Also we provide some Blender scripts so that you can visualize the curves and NerVE cube grid in Blender. Check utils/visualization_blender.

Dataset and Training

Firstly, a ready-to-use dataset can be downloaded for training:

NerVE64Dataset.zip

Note this dataset can be only used for NerVE grid of resolution 64^3. For preparing your dataset, check utils/prepare_data.

For training, the code can be simply run like:

python train_net.py -c /path/to/your/config.yaml

Some examples of the config file can be found in exp/template. Note that you should change root_path and root of dataset in the config file, before you start to train the model. Here we have three config files(cube.yaml, face.yaml, geom.yaml) for a complete NerVE model, because we found such separate training is easier to converge and enough to produce good results.

To evaluate the predicted curves, you can use functions from utils/pwl2CAD/eval_cad_curve.py to calculate their Chamfer distance and Hausdorff distance with ground truth curves.

More Results

Citation

If you find our work useful in your research, please consider citing:

@misc{zhu2023nerve,
  title={NerVE: Neural Volumetric Edges for Parametric Curve Extraction from Point Cloud}, 
  author={Xiangyu Zhu and Dong Du and Weikai Chen and Zhiyou Zhao and Yinyu Nie and Xiaoguang Han},
  year={2023},
  eprint={2303.16465},
  archivePrefix={arXiv},
  primaryClass={cs.CV}

}

uhzoaix / NerVE

readme