DGC-Net: Dense Geometric Correspondence Network

This is a PyTorch implementation of our work "DGC-Net: Dense Geometric Correspondence Network"

A CNN-based approach to obtain dense pixel correspondences between two views.

Installation

With conda

• create and activate conda environment with Python 3.x

conda create -n my_fancy_env python=3.7
source activate my_fancy_env

• install Pytorch v1.0.0 and torchvision library

pip install torch torchvision

• install all dependencies by running the following command:

pip install -r requirements.txt

With pip

• create and activate virtual environment with Python 3.x

python3 -m venv dgc-env
source dgc-env/bin/activate

• install Pytorch and torchvision library

pip install torch torchvision

• install all dependencies by running the following command:

pip install termcolor opencv-python tqdm pandas

Getting started

• eval.py demonstrates the results on the HPatches dataset To be able to run eval.py script:

  • Download an archive with pre-trained models click and extract it to the project folder
  • Download HPatches dataset (Full image sequences). The dataset is available here at the end of the page
  • Run the following command:

  python eval.py --image-data-path /path/to/hpatches-geometry
  

• train.py is a script to train DGC-Net/DGCM-Net model from scratch. To run this script, please follow the next procedure:

  • Download the TokyoTimeMachine dataset
  • Run the command:

  python train.py --image-data-path /path/to/TokyoTimeMachine
  

Performance on HPatches dataset

Method / HPatches ID	Viewpoint 1	Viewpoint 2	Viewpoint 3	Viewpoint 4	Viewpoint 5
PWC-Net	4.43	11.44	15.47	20.17	28.30
GM best model	9.59	18.55	21.15	27.83	35.19
DGC-Net (paper)	1.55	5.53	8.98	11.66	16.70
DGCM-Net (paper)	2.97	6.85	9.95	12.87	19.13
DGC-Net (repo)	1.74	5.88	9.07	12.14	16.50
DGCM-Net (repo)	2.33	5.62	9.55	11.59	16.48

Note: There is a difference in numbers presented in the original paper and obtained by the models of this repo. It might be related to the fact that both models (DGC-Net and DGCM-Net) have been trained using Pytorch v0.3.

More qualitative results are presented on the project page

License

Our code is released under the Creative Commons BY-NC-SA 3.0, available only for non-commercial use.

How to cite

If you use this software in your own research, please cite our publication:

@inproceedings{Melekhov+Tiulpin+Sattler+Pollefeys+Rahtu+Kannala:2018,
      title = {{DGC-Net}: Dense geometric correspondence network},
      author = {Melekhov, Iaroslav and Tiulpin, Aleksei and 
               Sattler, Torsten, and 
               Pollefeys, Marc and 
               Rahtu, Esa and Kannala, Juho},
       year = {2019},
       booktitle = {Proceedings of the IEEE Winter Conference on 
                    Applications of Computer Vision (WACV)}
}