MGCN

This code implements Multiscale Graph Convolutional Network (MGCN) in our SIGGRAPH 2020 paper:

"MGCN: Descriptor Learning using Multiscale GCNs"

by Yiqun Wang, Jing Ren, Dong-Ming Yan, Jianwei Guo, Xiaopeng Zhang, Peter Wonka.

Please consider citing the above paper if this code/program (or part of it) benefits your project.

Environment

	conda create -n MGCN python=3.7     # (options: 3.X)
	source activate MGCN                # (create and activate new environment if you use Anaconda)
	
	conda install pytorch=1.5.0 torchvision cudatoolkit=10.1 -c pytorch     # (options: 10.X)
	pip install torch-scatter==latest+cu101 -f https://pytorch-geometric.com/whl/torch-1.5.0.html
	pip install torch-sparse==latest+cu101 -f https://pytorch-geometric.com/whl/torch-1.5.0.html
	pip install torch-cluster==latest+cu101 -f https://pytorch-geometric.com/whl/torch-1.5.0.html
	pip install torch-geometric

Usage

Put models in the folder MGCN/datasets/Faust/data_mesh/ for FAUST dataset or MGCN/datasets/Scape/data_mesh/ for SCAPE dataset
generate Wavelet Energy Decomposition Signature (WEDS) descriptors and wavelets using Repository WEDS or directly download processed data (traning.pt and test.pt) below and put them into the folder MGCN/datasets/Faust/processed/ or MGCN/datasets/Scape/processed/.
Training examples

	# for FAUST
	python MGCN_FAUST.py
	# for SCAPE
	python MGCN_SCAPE.py

Restore checkpoint and generate descriptors

	# for FAUST
	python MGCN_FAUST.py --uc --gd -l --ln=mgcn_faust-300
	# for SCAPE
	python MGCN_SCAPE.py --uc --gd -l --ln=mgcn_scape-300

Dense correspondence can be obtained by directly computing the nearest-neighbor using the L2 distance in descriptor space, and this Repository can be used to visualize the descriptors.

Processed data

Processed Dataset	Download Link	Description
FAUST original	Google Drive, 1G	75 models for training and 15 models for testing (6890 points)
FAUST 5 resolutions	Google Drive, 1G	15 X 5 models for testing (6890, 8K, 10K, 12K, 15K points)
SCAPE remeshed	Google Drive, 623M	61 models for training and 10 models for testing (~5K points)
SCAPE original	Google Drive, 218M	10 models for testing (12.5K points)

Cite

@article{wang2020mgcn,
  title=      {MGCN: Descriptor Learning using Multiscale GCNs},
  author=     {Wang, Yiqun and Ren, Jing and Yan, Dong-Ming and Guo, Jianwei and Zhang, Xiaopeng and Wonka, Peter},
  journal=    {ACM  Trans. on Graphics (Proc. {SIGGRAPH})},
  year=       {2020},
}

License

This program is free software; you can redistribute it and/or modify it under the terms of the GNU General Public License as published by the Free Software Foundation; either version 2 of the License, or (at your option) any later version.

Chao-Jiang / MGCN