===== HF-Net: Robust Hierarchical Localization at Large Scale =====

We introduce a 6-DoF visual localization method that is accurate, scalable, and efficient, using HF-Net, a monolithic deep neural network for descriptor extraction. The proposed solution achieves state-of-the-art accuracy on several large-scale public benchmarks while running in real-time.

[[https://arxiv.org/abs/1812.03506|arXiv paper link]]

{{:hfnet_teaser.jpg?nolink&400|}}

==== Download SfM Models and Trained Weights ====

Navigate to http://robotics.ethz.ch/~asl-datasets/2019_CVPR_hierarchical_localization/ for the list of all files available for download or select specific items in the list below.

== SfM Models ==
SfM models built with SuperPoint, and usable with HF-Net, for the following datasets: Aachen Day-Night, CMU Seasons, Extended CMU Seasons and RobotCar Seasons.

[[http://robotics.ethz.ch/~asl-datasets/2019_CVPR_hierarchical_localization/sfm_aachen.tar.gz|Aachen Day-Night]]

[[http://robotics.ethz.ch/~asl-datasets/2019_CVPR_hierarchical_localization/sfm_cmu.tar.gz|CMU Seasons]]

[[http://robotics.ethz.ch/~asl-datasets/2019_CVPR_hierarchical_localization/sfm_cmu_extended.tar.gz|Extended CMU Seasons]]

[[http://robotics.ethz.ch/~asl-datasets/2019_CVPR_hierarchical_localization/sfm_robotcar.tar.gz|RobotCar Seasons]]

== Trained Weights and TF Graph for HF-Net ==

The Tensorflow graph and trained weights.

[[http://robotics.ethz.ch/~asl-datasets/2019_CVPR_hierarchical_localization/hfnet_tf.tar.gz|
HF-Net Tensorflow]]


==== GitHub Repository ====

Apart from this collection of SfM models and the trained Tensorflow graph, we also share our source code to train the network, perform inference and full 6-DoF pose retrieval.

[[https://github.com/ethz-asl/hfnet|Go to Github]]

The soures in https://github.com/ethz-asl/hfnet allow you to:
  * Perform state-of-the-art 6-DoF hierarchical localization using a flexible Python pipeline
  * Train HF-Net with multi-task distillation in TensorFlow
  * Evaluate feature detectors and descriptors on standard benchmarks
  * Build Structure-from-Motion models based on state-of-the-art learned features

==== Citation ====

Please consider citing the corresponding publication if you use this work in an academic context:
<code>
@inproceedings{sarlin2019coarse,
  title={From Coarse to Fine: Robust Hierarchical Localization at Large Scale},
  author={Sarlin, P.-E. and Cadena, C. and Siegwart, R. and Dymczyk, M.},
  article={CVPR},
  year={2019}
}
</code>