GCN-Denoiser: Mesh Denoising with Graph Convolutional Networks

Qt and Pytorch implementation for our paper "GCN-Denoiser: Mesh Denoising with Graph Convolutional Networks" (ACM Transactions on Graphics 2022)

We propose GCN-Denoiser, a novel feature-preserving mesh denoising method based on graph convolutional networks (GCNs). Unlike previous learning-based mesh denoising methods that exploit hand-crafted or voxel-based representations for feature learning, our method explores the structure of a triangular mesh itself and introduces a graph representation followed by graph convolution operations in the dual space of triangles. We also create a new dataset called PrintData containing 20 real scans with their corresponding ground truths for the research community.

Denoised Results:

Interface:

Code:

Prerequisites:

• Hardware: Personal computer with NVIDIA GPU.
• Environments: CUDA10.0, Windows system (network training part can also be used on Linux).

Third Party Library:

• Pytroch C++ 1.2.0 , Eigen, Flann and OpenMesh at runtime.
• Pytorch 1.2.0, numpy, Scipy 1.4.1 and tensorbordx 1.13 (>python3.5) in training stage.

Network part:

The training code and part of validation data are supplied. Network tests can be run by:

cd DenoisingGCN/testSamples
unzip bunny_0_2.zip
cd ../
python datautils.py
python test.py

bunny_0_2/*.mat are sampled patches from the noisy bunny model with 0.2 level of Gaussian noise.

Denoising Interface:

Executable demo, the corresponding code, and some sampled meshes are supplied.

• For .exe, windows platform is required and the CUDA PATH must be set in the system environment. Some .dll are required (CUDA&LibTorch: c10.dll, c10_cuda.dll, caffe2_nvrtc.dll, nvToolsExt61_1.dll, torch.dll; Qt: Qt5Core.dll, Qt5Gui.dll, Qt5OpenGL.dll, Qt5Widgets.dll). Required dlls can be downloaded from our Google Drive link.

• For code, Visual Studio 2017 and Qt 5.12 are required.

Pre-trained models:

One version of GCN pre-trained model for synthetic models is supplied.

Results:

Our results of public noisy meshes proposed by CNR are released in our Google Drive link (including synthetic CAD, smooth and featured models and Kinect scanned models).

Dataset:

Download the dataset from our Google Drive link. Our denoised results and evaluation code can also be found in this Google Drive link.

Citation

If you find this useful for your research, please cite the following paper.

@article{shen2022gcndenoiser,
  title={GCN-Denoiser: Mesh Denoising with Graph Convolutional Networks},
  author={Shen, yuefan and Fu, Hongbo and Du, Zhongshuo and Chen, Xiang and Burnaev, Evgeny and Zorin, Denis and Zhou, Kun and Zheng, Youyi},
  journal={ACM Trans. Graph.},
  volume={41},
  number={1},
  issn={0730-0301},
  numpages={14},
  year={2022}
}

Waiting for updating...

Acknowledgements

Part of this implementation is based on DGCNN and GNF.

Our paper was originally submitted in May 2020, and after a long journey, it was finally accepted in August 2021 and published in 2022. We would like to express our gratitude to all the anonymous reviewers for their constructive comments.