MMNet

This repo is the official implementation of ICCV 2021 paper "Multi-scale Matching Networks for Semantic Correspondence.".

Pre-requisite

conda create -n mmnet python==3.8.0
conda activate mmnet
conda install torch==1.8.1 torchvision==0.9.1
pip install matplotlib scikit-image pandas

for installation of gluoncvth (fcn-resnet101):

git clone https://github.com/StacyYang/gluoncv-torch.git
cd gluoncv-torch
python setup.py install

Reproduction

for training

python train.py --seed 0 --lr 0.0005

for test

Trained models are available on [google drive].

pascal with fcn-resnet101 backbone(PCK@0.05:81.6%):

python test.py --alpha 0.05 --backbone fcn-resnet101 --ckp_name path\to\ckp_pascal_fcnres101.pth --resize 224,320

spair with fcn-resnet101 backbone(PCK@0.1):

python test.py --alpha 0.05 --benchmark spair --backbone fcn-resnet101 --ckp_name path\to\ckp_spair_fcnres101.pth --resize 224,320

Bibtex

If you use this code for your research, please consider citing:

@article{zhao2021multi,
  title={Multi-scale Matching Networks for Semantic Correspondence},
  author={Zhao, Dongyang and Song, Ziyang and Ji, Zhenghao and Zhao, Gangming and Ge, Weifeng and Yu, Yizhou},
  booktitle={Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV)},
  year={2021}
}