Awesome

Dual Deep Mesh Prior [ECCV2022]

Paper | Supp | Poster

The official implementation of Learning Self-prior for Mesh Denoising using Dual Graph Convolutional Networks, ECCV2022.

A deep-learning framework for mesh denoising from a single noisy input, where two graph convolutional networks are trained jointly to filter vertex positions and facet normals apart.

Method Overview

Results

Getting Started

0. Environments

python==3.10
torch==1.13.1
torch-geometric==2.2.0

1. Installation

git clone https://github.com/astaka-pe/Dual-DMP
cd Dual-DMP
docker image build -t astaka-pe/ddmp .
docker run -itd --gpus all -p 8080:8080 --name ddmp -v .:/work astaka-pe/ddmp
docker exec -it ddmp /bin/bash

2. Preparation

The Dataset is distributed as a zip file. Please unzip and place it under Dual-DMP directory.

unzip datasets.zip

3. Training

• CAD model

python3 main.py -i datasets/fandisk --k1 3 --k2 0 --k3 3 --k4 4 --k5 2 --bnfloop 5

• Non-CAD model

python3 main.py -i datasets/ankylosaurus

• Real-scanned model

python3 main.py -i datasets/pyramid --iter 50

You can monitor the training progress through the web viewer. (Default: http://localhost:8080)

Outputs will be generated under datasets/{model-name}/output/ with their MAD scores.

Appendix

Training with your own data

Place a noisy mesh and a ground-truth mesh under datasets/{model-name}/ .

• Noisy mesh: {model-name}_noise.obj
• Ground-truth mesh: {model-name}_gt.obj

Run

python3 preprocess/preprocess.py -i datasets/{model-name}

for edge-based normalization and creating initial smoothed mesh.

Finally, run

python3 main.py -i datasets/{model-name}

You should set appropriate weights as discribed in the paper.

Training without using ground-truth data

After runnning preprocess.py, run

python3 main4real.py -i datasets/{model-name}

Creating noisy data

Run

python3 preprocess/noisemaker.py -i datasets/{model-name}/{model-name}.obj --level {noise-level}

Citation

@InProceedings{hattori2022ddmp,
  author        = {Hattori, Shota and Yatagawa, Tatsuya and Ohtake, Yutaka and Suzuki, Hiromasa},
  title         = {Learning Self-prior for Mesh Denoising using Dual Graph Convolutional Networks},
  booktitle     = {Proceedings of the European Conference on Computer Vision (ECCV)},
  year          = {2022},
  doi           = {10.1007/978-3-031-20062-5_21}
}