joonahn / hpf

Official PyTorch Implementation of Hyperpixel Flow, ICCV 2019

Hyperpixel Flow:
Semantic Correspondence with Multi-layer Neural Features

This is the implementation of the paper "Hyperpixel Flow: Semantic Correspondence with Multi-layer Neural Features" by J. Min, J. Lee, J. Ponce and M. Cho. Implemented on Python 3.6 and Pytorch 1.0.1.

For more information, check out project [website] and the paper on [arXiv].

Conda environment settings

conda create -n hpf python=3.6
conda activate hpf

cat /usr/local/cuda/version.txt
conda install pytorch=1.0.1 torchvision cudatoolkit=10.0 -c pytorch (if CUDA 10) 
conda install pytorch=1.0.1 torchvision cudatoolkit=9.0 -c pytorch (if CUDA 9) 

conda install -c anaconda scikit-image
conda install -c anaconda pandas
conda install -c anaconda requests
conda install pillow=6.1
pip install gluoncv-torch

Reproduction

Beam search on SPair-71k validation set:

python beamsearch.py --dataset spair --thres bbox --backbone resnet50
python beamsearch.py --dataset spair --thres bbox --backbone resnet101

Beam search on PF-PASCAL validation set:

python beamsearch.py --dataset pfpascal --thres bbox --backbone resnet50
python beamsearch.py --dataset pfpascal --thres bbox --backbone resnet101  

Results on PF-PASCAL: (PCK: 83.4%, 84.8%, 88.3%)

python evaluate.py --dataset pfpascal --backbone resnet50 --hyperpixel '(2,7,11,12,13)'
python evaluate.py --dataset pfpascal --backbone resnet101 --hyperpixel '(2,17,21,22,25,26,28)'
python evaluate.py --dataset pfpascal --backbone fcn101 --hyperpixel '(2,4,5,18,19,20,24,32)'

Results on PF-WILLOW: (PCK: 74.4%)

python evaluate.py --dataset pfwillow --backbone resnet101 --hyperpixel '(2,17,21,22,25,26,28)'

Results on Caltech-101: (LT-ACC: 0.88, IoU: 0.64)

python evaluate.py --dataset caltech --backbone resnet50 --hyperpixel '(2,7,11,12,13)'

Results on SPair-71k: (PCK: 27.2%, 28.2%)

python evaluate.py --dataset spair --backbone resnet50 --hyperpixel '(0,9,10,11,12,13)'
python evaluate.py --dataset spair --backbone resnet101 --hyperpixel '(0,8,20,21,26,28,29,30)'

To visualize predictions using TPS transformation, add command line argument --visualize:

python evaluate.py --visualize

Bibtex

If you use this code and SPair-71k dataset for your research, please consider citing:

@InProceedings{min2019hyperpixel, 
   title={Hyperpixel Flow: Semantic Correspondence with Multi-layer Neural Features},
   author={Juhong Min and Jongmin Lee and Jean Ponce and Minsu Cho},
   booktitle={ICCV},
   year={2019}
}

@article{min2019spair,
   title={SPair-71k: A Large-scale Benchmark for Semantic Correspondence},
   author={Juhong Min and Jongmin Lee and Jean Ponce and Minsu Cho},
   journal={arXiv prepreint arXiv:1908.10543},
   year={2019}
}