HF-Net: Robust Hierarchical Localization at Large Scale

Year: 2019
DOI: 10.3929/ethz-c-000788517

Structure-from-Motion models and trained HF-Net weights for large-scale 6-DoF localization. HF-Net is a monolithic network for feature detection/description enabling accurate, scalable, real-time localization.

Paper: arXiv • Code: hfnet

Contents

• SfM models built with SuperPoint (usable with HF-Net) for:
  • Aachen Day-Night
  • CMU Seasons
  • Extended CMU Seasons
  • RobotCar Seasons
• Trained TensorFlow graph and weights for HF-Net

Downloads

• Dataset directory: https://doi.org/10.3929/ethz-c-000788517 (was 2019_CVPR_hierarchical_localization/)
• SfM models:
  • https://doi.org/10.3929/ethz-c-000788517 (was sfm_aachen.tar.gz)
  • https://doi.org/10.3929/ethz-c-000788517 (was sfm_cmu.tar.gz)
  • https://doi.org/10.3929/ethz-c-000788517 (was sfm_cmu_extended.tar.gz)
  • https://doi.org/10.3929/ethz-c-000788517 (was sfm_robotcar.tar.gz)
• HF-Net TensorFlow graph + weights: https://doi.org/10.3929/ethz-c-000788517 (was hfnet_tf.tar.gz)

Citation

P.-E. Sarlin, C. Cadena, R. Siegwart, M. Dymczyk,
“From Coarse to Fine: Robust Hierarchical Localization at Large Scale”, CVPR 2019.

Tags: localization, sfm, features, vision, 6dof