GS-Net: Point Cloud Sampling with Graph Neural Networks

Environment

Environment:Python3.6.9, PyTorch1.7.1, gcc 6 or 7 or higher!(gcc 4 will get ninja bug, this is important)

if meet ninja bug

PW: if you get bug like: "RuntimeError: Ninja is required to load C++ extensions" "Command ‘[‘ninja‘, ‘-v‘]‘ returned non-zero exit status 1" may be the problem is ninja version or ninja install error, you should reinstall ninja.

renew debug method: there are 2 methods can resolve this problem. The first one is change "command = ['ninja', '-v']" to "command = ['ninja', '--version']" in "envs/gsnet/lib/python3.6/site-packages/torch/utils/cpp_extension.py/_run_ninja_build". The seconde one is reinstall ninja. I recommend the first method.

Reinstall ninja:
First: install re2c in http://re2c.org/index.html
Here I install re2c3.0: wget https://github.com/skvadrik/re2c/releases/download/3.0/re2c-3.0.tar.xz

Second: unzip and make install re2c-3.0

tar -xvf re2c-3.0.tar.xz 
cd re2c-3.0 
autoreconf -i -W all 
./configure --prefix=/home/yourName/ninjatest 
make 
make install 
make check   (for test re2c install) 

Third: export re2c PATH

vi ~/.bashrc (and add low command) 
export PATH="/disk2/cxl/ninjatest/bin:$PATH" 
source ~/.bashrc 
re2c -V (test re2c) 

Fourth: start install ninja

cd .. 
git clone https://github.com/ninja-build/ninja.git && cd ninja 
./configure.py --bootstrap 

fifth: export ninja PATH

vi ~/.bashrc (and add low command)
export PATH="/disk2/cxl/ninja:$PATH" 
source ~/.bashrc 
ninja --version  

Install environment

1.install pointnet2_ops

down load Pointnet2_Pytorch:
git clone https://github.com/erikwijmans/Pointnet2_PyTorch.git

then install CUDA kernels(use grouping_operation in softprojection in samplenet)

pip install pointnet2_ops_lib/.

2.install KNN_CUDA KNN_CUDA provide by https://github.com/unlimblue/KNN_CUDA

First: install KNN_CUDA

pip install --upgrade https://github.com/unlimblue/KNN_CUDA/releases/download/0.2/KNN_CUDA-0.2-py3-none-any.whl

Second: import knn_cuda in python3 (in terminal), may be post error "RuntimeError: Ninja is required to load C++ extensions" https://blog.csdn.net/xiaoyaolangwj/article/details/119382717

3.install PyG(torch_geometric) for building Point Graph https://github.com/pyg-team/pytorch_geometric https://zhuanlan.zhihu.com/p/381204915

actually, you can install by following command, （与1.7.0相同版本）:

pip install https://data.pyg.org/whl/torch-1.7.0%2Bcu102/torch_scatter-2.0.7-cp36-cp36m-linux_x86_64.whl
pip install https://data.pyg.org/whl/torch-1.7.0%2Bcu102/torch_sparse-0.6.9-cp36-cp36m-linux_x86_64.whl
pip install https://data.pyg.org/whl/torch-1.7.0%2Bcu102/torch_cluster-1.5.9-cp36-cp36m-linux_x86_64.whl
pip install https://data.pyg.org/whl/torch-1.7.0%2Bcu102/torch_spline_conv-1.2.1-cp36-cp36m-linux_x86_64.whl
pip install torch-geometric

dataset prepare

ModelNet40: https://modelnet.cs.princeton.edu/
ScanObjectNN: https://hkust-vgd.github.io/scanobjectnn/
SemanticKITTI-cls: https://drive.google.com/file/d/1mf0GIVmd0BU_lakTBGpMcdT1jrCxdUyM/view?usp=sharing

data
├── modelnet40_ply_hdf5_2048 # ModelNet40
├── scanobjectnn/h5_files/main_split # ScanObjectNN
└── semantickitti_cls # SemanticKITTI-cls

semanticKITTI-cls construction

SemanticKITTI-cls, which is a custom-built dataset based on the LiDAR dataset SemanticKITTI, containing point cloud objects from outdoor driving scenes. SemanticKITTI-cls consists of 4,000 samples across 6 categories. As objects move farther from the sensor, these point clouds exhibit sparsity and diverse density characteristics.

Here is the SemanticKITTI-cls's distribution.

Split	Car	Motorcycle	Truck	Other-vehicle	Person	Bicyclist	Total
Train	1000	200	600	1000	80	100	2980
Test	300	87	254	300	34	45	1020
Total	1300	287	854	1300	114	145	4000

Usage

Classifier Training:

train pointnet classifier on Modelnet40 use:

 python train_classifier --dataset modelnet40 --num_category 40 

train pointnet classifier on ScanObjectNN use:

 python train_classifier --dataset scanobjectnn --dataclass OBJ_BG --num_category 15

train pointnet classifier on SemanticKITTI-cls use:

 python train_classifier --dataset semantickitti_cls --num_category 6

Sampling Network Training:

After train classifier, joinly train sampling network. For training GS-Net on ModelNet40 and PointNet classifier with sampling rate r=32, you can use:

 python train_gsnet.py --dataset modelnet40 --classifier_model pointnet_cls --classifier_model_path weights/your_PointNet_classifier_modelnet_path.pth --num_category 40 --not_debug --assign_ratio 0.03125

For training GS-Net on ScanObjectNN and PointNet classifier with sampling rate r=32, you can use:

 python train_gsnet.py --dataset scanobjectnn --classifier_model pointnet_cls --classifier_model_path weights/your_PointNet_classifier_scanobjectnn_path.pth --num_category 15 --not_debug --assign_ratio 0.03125

For training GS-Net on SemanticKITTI-cls and PointNet classifier with sampling rate r=32, you can use:

 python train_gsnet.py --dataset semantickitti_cls --classifier_model pointnet_cls --classifier_model_path weights/your_PointNet_classifier_semantickitticls_path.pth --num_category 6 --not_debug --assign_ratio 0.03125 --alpha 0.01

Classifier Evaluating:

if need visulization add --vis_sampled_point, if need save to ply add --save_sampled_point evaluate pointnet classifier on ModelNet40 use:

python eval_classifier.py --dataset modelnet40 --classifier_model_path weights/PointNet_classifier_model_modelnet.pth  --num_category 40 --not_debug 

evaluate pointnet classifier on ScanObjectNN use:

python eval_classifier.py --dataset scanobjectnn --classifier_model_path weights/PointNet_classifier_model_scanobjectnn.pth  --num_category 15 --not_debug 

evaluate pointnet classifier on SemanticKITTI-cls use:

python eval_classifier.py --dataset semantickitti_cls --classifier_model_path weights/PointNet_classifier_model_semantickitti_cls.pth  --num_category 6 --not_debug 

Sampling network Evaluating:

Evaluating GS-Net base on PointNet classifier and ModelNet40 with r=32, you can use:

python eval_gsnet.py --dataset modelnet40 --not_debug --match --assign_ratio 0.03125 --points_noise 0.1 --sampler_model_path log/gsnet_pyg/log_dir/checkpoints/best_model.pth

Retrieval Evaluating:

You need to add --save_retrieval_vectors for retrieval evaluating:

python eval_gsnet.py --dataset modelnet40 --not_debug --match --assign_ratio 0.03125 --points_noise 0.1 --sampler_model_path log/gsnet_pyg/log_dir/checkpoints/best_model.pth --save_retrieval_vectors

python analyze_precision_recall.py --dump_dir log_eval/gsnet_pyg/log_dir_32/retrieval/ --num_out_points 32 --model_name GSNet

Citation

If you find GS-Net useful for your research and applications, please cite using this BibTeX:

@article{chen2025gs,
  title={GS-Net: Point cloud sampling with graph neural networks},
  author={Chen, Xiaolei and Chen, Jie and Qiu, Shoumeng and Xue, Xiangyang and Pu, Jian},
  journal={Pattern Recognition},
  pages={112054},
  year={2025},
  publisher={Elsevier}
}

Thanks

Thanks for Samplenet, Pointnet_Pytorch,PointNet_Pytorch(for environment install).