Skip to content


Repository files navigation

Deep Graphical Feature Learning for the Feature Matching Problem, ICCV2019

Created by Zhen Zhang and Wee Sun Lee.


If you find the code useful, please consider citing

    Author = {Zhen Zhang and Wee Sun Lee},
    Title = {Deep Graphical Feature Learning for the Feature Matching Problem},
    Year = {2019},
    booktitle = {Proceedings of the IEEE International Conference on Computer Vision},


Please install the following dependencies for training and testing

conda create -n python3.6 python=3.6
conda activate python3.6
conda install tensorflow
conda install conda install pytorch torchvision cudatoolkit=10.0 -c pytorch # adjust the cuda version according to your platform
conda install scikit-image
pip install tqdm

Train the model

The following code can be used to train the model:

python --batch_size 64 # on RTX2080Ti, uses about 9GB GPU memory

Test the model

After training over random generated 9M samples, the training code will finally generate a parameter file ``''. As in our code the random matching pairs are generated on the fly, it is equivalent to training over 9M samples for one epoch.

Synthetic data

To reproduce the experimental results on synthetic data, please run the following script:

python --param_path ./

CMU House

The original link to download the CMU house dataset is not valid anymore, thus the data is included in the repo. To reproduce the experimental results on the dataset, please run the script as follows

python --param_path ./ --data_path ./datasets/cmum/house


To reproduce the experimental results on PF-Pascal dataset, please first download the PF-Pascal dataset by running the script as follows,

pushd datasets

After that, the results can be reproduces by running the following script

python --param_path ./ --data_path ./datasets/PF-dataset-PASCAL/ --random_rotate False 
python --param_path ./ --data_path ./datasets/PF-dataset-PASCAL/ --random_rotate True


No description, website, or topics provided.







No releases published


No packages published