GN-Net Benchmark

For more information see https://vision.in.tum.de/gn-net

This code allows to generate pixel-correspondences for the GN-Net benchmark, presented in

"GN-Net: The Gauss-Newton Loss for Multi-Weather Relocalization".
L. von Stumberg, P. Wenzel, Q. Khan and D. Cremers. RA-L 2020

Install Dependencies

pip install opencv-python
pip install scipy==1.1.0
pip install scikit-image

Running

See example.py how to generate matches. At the moment matches can only be generated for the Carla sequences. For generating the matches either the groundtruth-depths or depths exported from Stereo DSO can be used. To specify this you can adjust the parameter use_dso_depths of the generate_matches_carla

For a start you can simply run

python example.py