This repository contains pytorch implementation for RefineNet: Refine-Net: Normal Refinement Neural Network for Noisy Point Clouds (TPAMI'2022).

RefineNet presents a normal refinement network for estimating normals in noisy point clouds. This repository contains our network implementation and MPFS initial normal estimation. RefineNet can also be applied on other methods and refine their predicted normals. If you find our work useful in your research, please cite our paper:

@article{zhou2022refine,
    title={Refine-Net: Normal Refinement Neural Network for Noisy Point Clouds},
    author={Zhou, Haoran and Chen, Honghua and Zhang, Yingkui and Wei, Mingqiang and Xie, Haoran and Wang, Jun and Lu, Tong and Qin, Jing and Zhang, Xiao-Ping},
    journal={IEEE Transactions on Pattern Analysis and Machine Intelligence},
    year={2022},
    publisher={IEEE}
}

• The code has been tested on this configuration:

  • PyTorch == 1.1.0
  • python == 3.7.10
  • CUDA == 10.1
  • scikit-learn

• First, you should compile the C++ extension modules for python located in cpp_wrappers. Open a terminal in this folder, and run:

    sh compile_wrappers.sh
  

The original pcpnet dataset can be downloaded from here. We also provide our processed dataset and normal results here.

Our synthetic benchmark from here.

Besides, we provide other RefineNet results on Kinect_v1 and NYU_V2.

Before training, specify your dataset directories in config_pcpnet.py:

cfg.dataset.pointcloud_dir = '<*PATH-TO-YOUR-DATASET*>/pcpnet'

RefineNet takes a set of initial normals as input. You can use your own initial normal results on the training set. Alternatively, we provide our normal input here, or you can use the processed data here (with processed features). Place any of these data under:

cfg.dataset.normal_dir = '<*PATH-TO-NORMALS*>/initial_normals'

To train RefineNet, simply run:

sh train_pcpnet.sh

To test RefineNet, you should provide the pretrained models:

python3 train.py --config pcpnet --test --pretrained <*PATH-TO-PRETRAINED*>

For example, the default path for pretrained models is ./results/train_Config_pcpnet:

python3 train.py --config pcpnet --test --pretrained train_Config_pcpnet

To use our synthetic dataset for training, specify the directories in config_synthetic.py:

cfg.dataset.pointcloud_dir = '<*PATH-TO-YOUR-DATASET*>/benchmark'
cfg.dataset.normal_dir = '<*PATH-TO-NORMALS*>/initial_normals'

Then, run:

sh train_synthetic.sh

Similarly, to test a pretrained model:

python3 train.py --config synthetic --test --pretrained <*PATH-TO-PRETRAINED*>

Our initial normal estimator (MFPS) is implemented in ./matlab. Simply run ./matlab/RUN.m to compute initial normals for each input point cloud. Matlab >= 2018b is required.

To use your own dataset, simply change the data path in ./matlab/RUN.m:

path_pointcloud = '<*PATH-TO-YOUR-DATASET*>';

• 2022-03-18: Initial Update.
• 2022-04-08: Update synthetic benchmark and other results.

This work is partly based on our previous cvpr20 paper.

Some of the code is borrowed from pcpnet. We thank the authors for their excellent work.