Uploaded April 2026 | Updated November 2026, 5 days ago
In this work, we tackle the challenge of ground-to-aerial global LiDAR localisation, without access to GPS or SLAM as prior information! Our method, HOTFLoc++, extracts robust hierarchical features from dense forest point clouds with our octree-based HOTFormerLoc backbone, which we train end-to-end for place recognition, re-ranking, and 6-DoF metric localisation. Our learnable multi-scale geometric verification module adaptively re-ranks place recognition retrievals by analysing geometric consistency at multiple granularities to improve robustness in cluttered scenes. We then employ a coarse-to-fine metric localisation approach that estimates the full 6-DoF pose via efficient patch-wise registration.
Our extensive experiments demonstrate the superiority of our approach for ground-to-aerial global localisation in forests, with ~30% higher Recall@1 than SOTA, and ~86% of queries localised within 2m and 5° of ground truth (a ~2x improvement over SOTA!)
🔗 Paper (arXiv): lnkd.in/gMPshkGz
💻 Code: github.com/csiro-robotics/HOTFLoc
In this work, we tackle the challenge of ground-to-aerial global LiDAR localisation, without access to GPS or SLAM as prior information! Our method, HOTFLoc++, extracts robust hierarchical features from dense forest point clouds with our octree-based HOTFormerLoc backbone, which we train end-to-end for place recognition, re-ranking, and 6-DoF metric localisation. Our learnable multi-scale geometric verification module adaptively re-ranks place recognition retrievals by analysing geometric consistency at multiple granularities to improve robustness in cluttered scenes. We then employ a coarse-to-fine metric localisation approach that estimates the full 6-DoF pose via efficient patch-wise registration.
Our extensive experiments demonstrate the superiority of our approach for ground-to-aerial global localisation in forests, with ~30% higher Recall@1 than SOTA, and ~86% of queries localised within 2m and 5° of ground truth (a ~2x improvement over SOTA!)
🔗 Paper (arXiv): lnkd.in/gMPshkGz
💻 Code: github.com/csiro-robotics/HOTFLoc







![Our facilities [shorter] - Robotics and Autonomous Systems Group CSIROs Data61
We work with a broad range of industries, and are open for collaboration.
Want to partner with us or use our research facilities? Contact us to learn more.
https://research.csiro.au/robotics/who-we-are/our-facilities/ Our facilities [shorter] - Robotics and Autonomous Systems Group CSIROs Data61](https://i.ytimg.com/vi/MbvtDm8ebC4/mqdefault.jpg)


