DGC-Net: Dense Geometric Correspondence Network

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

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

Installation

• 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

Getting started

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

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

• 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:

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

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

• Download the TokyoTimeMachine dataset

• Run the command:

Performance on HPatches dataset

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

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)}
}