Skip to content
No description, website, or topics provided.
Python Shell
Branch: master
Clone or download
Fetching latest commit…
Cannot retrieve the latest commit at this time.
Type Name Latest commit message Commit time
Failed to load latest commit information.

Deep Fundamental Matrix Estimation

This repository contains code for training a network as described in our paper:

Deep Fundamental Matrix Estimation
René Ranftl and Vladlen Koltun
European Conference on Computer Vision (ECCV), 2018

The trained network can be used to compute a fundamental matrix based on noisy point correspondences.


Create and activate conda environment:

conda env create -f environment.yml
conda activate dfe


Training can be performed with the following command:

python --dataset [dataset folders]

Dataset folders are assumed to be in COLMAP format. Multiple folders can be listed by separating them with a whitespace.

An example dataset can be found here: Family.

To get help:

python -h


python --dataset [dataset folders] --model [path to model]

A pre-trained model that was trained on Tanks and Temples can be found in the models folder.

Creating your own dataset

The dataloader assumes that a dataset folder contains the following files:


cameras.bin and images.bin are SfM reconstruction in COLMAP format.

reconstruction.db contains descriptors, matches, etc.

These files are automatically created when performing reconstruction using COLMAP. We found that better results are achieved with less aggressive filtering for feature matching than the default COLMAP settings. The recommended way to create a training set is:

  1. Run COLMAP to produce cameras.bin and images.bin and initial reconstruction.db
  2. Run to produce a training set with higher outlier ratios to replace reconstruction.db


Please cite our paper if you use this code in your research:

    author = {Ranftl, Rene and Koltun, Vladlen},
    title = {Deep Fundamental Matrix Estimation},
    booktitle = {The European Conference on Computer Vision (ECCV)},
    year = {2018}


MIT License

You can’t perform that action at this time.