Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

HOTFLoc++: End-to-End Hierarchical LiDAR Place Recognition, Re-Ranking, and 6-DoF Metric Localisation in Forests

Ethan Griffiths, Maryam Haghighat, Simon Denman, Clinton Fookes, Milad Ramezani

TL;DR

HOTFLoc++ tackles LiDAR place recognition and 6-DoF metric localisation in forests by combining an octree-based transformer backbone for hierarchical, multi-scale features with a learnable Multi-Scale Geometric Verification re-ranking module and a coarse-to-fine, patch-based registration head. The system is trained end-to-end with losses for place recognition, re-ranking, and localisation, and demonstrates strong Recall@1 improvements and robust cross-source performance on CS-Wild-Places, Wild-Places, and MulRan, while delivering two orders of magnitude faster registration than RANSAC. The MSGV re-ranking outperforms single-scale SGV by effectively aggregating information across scales, addressing aliasing and degraded features common in natural environments. The results indicate practical viability for real-time, long-term autonomous navigation in GPS-denied, unstructured outdoors, with a scalable architecture that supports future multi-modality integration.

Abstract

This article presents HOTFLoc++, an end-to-end framework for LiDAR place recognition, re-ranking, and 6-DoF metric localisation in forests. Leveraging an octree-based transformer, our approach extracts hierarchical local descriptors at multiple granularities to increase robustness to clutter, self-similarity, and viewpoint changes in challenging scenarios, including ground-to-ground and ground-to-aerial in forest and urban environments. We propose a learnable multi-scale geometric verification module to reduce re-ranking failures in the presence of degraded single-scale correspondences. Our coarse-to-fine registration approach achieves comparable or lower localisation errors to baselines, with runtime improvements of two orders of magnitude over RANSAC for dense point clouds. Experimental results on public datasets show the superiority of our approach compared to state-of-the-art methods, achieving an average Recall@1 of 90.7% on CS-Wild-Places: an improvement of 29.6 percentage points over baselines, while maintaining high performance on single-source benchmarks with an average Recall@1 of 91.7% and 96.0% on Wild-Places and MulRan, respectively. Our method achieves under 2 m and 5 degrees error for 97.2% of 6-DoF registration attempts, with our multi-scale re-ranking module reducing localisation errors by ~2$\times$ on average. The code will be available upon acceptance.

HOTFLoc++: End-to-End Hierarchical LiDAR Place Recognition, Re-Ranking, and 6-DoF Metric Localisation in Forests

TL;DR

HOTFLoc++ tackles LiDAR place recognition and 6-DoF metric localisation in forests by combining an octree-based transformer backbone for hierarchical, multi-scale features with a learnable Multi-Scale Geometric Verification re-ranking module and a coarse-to-fine, patch-based registration head. The system is trained end-to-end with losses for place recognition, re-ranking, and localisation, and demonstrates strong Recall@1 improvements and robust cross-source performance on CS-Wild-Places, Wild-Places, and MulRan, while delivering two orders of magnitude faster registration than RANSAC. The MSGV re-ranking outperforms single-scale SGV by effectively aggregating information across scales, addressing aliasing and degraded features common in natural environments. The results indicate practical viability for real-time, long-term autonomous navigation in GPS-denied, unstructured outdoors, with a scalable architecture that supports future multi-modality integration.

Abstract

This article presents HOTFLoc++, an end-to-end framework for LiDAR place recognition, re-ranking, and 6-DoF metric localisation in forests. Leveraging an octree-based transformer, our approach extracts hierarchical local descriptors at multiple granularities to increase robustness to clutter, self-similarity, and viewpoint changes in challenging scenarios, including ground-to-ground and ground-to-aerial in forest and urban environments. We propose a learnable multi-scale geometric verification module to reduce re-ranking failures in the presence of degraded single-scale correspondences. Our coarse-to-fine registration approach achieves comparable or lower localisation errors to baselines, with runtime improvements of two orders of magnitude over RANSAC for dense point clouds. Experimental results on public datasets show the superiority of our approach compared to state-of-the-art methods, achieving an average Recall@1 of 90.7% on CS-Wild-Places: an improvement of 29.6 percentage points over baselines, while maintaining high performance on single-source benchmarks with an average Recall@1 of 91.7% and 96.0% on Wild-Places and MulRan, respectively. Our method achieves under 2 m and 5 degrees error for 97.2% of 6-DoF registration attempts, with our multi-scale re-ranking module reducing localisation errors by ~2 on average. The code will be available upon acceptance.

Paper Structure

This paper contains 17 sections, 8 equations, 2 figures, 9 tables.

Table of Contents

  1. Introduction
  2. Related Works
  3. LiDAR Place Recognition
  4. Metric Localisation
  5. Re-Ranking
  6. Method
  7. Place Recognition with Hierarchical Features
  8. Learnable Multi-Scale Geometric Verification
  9. Coarse-to-Fine Registration
  10. End-to-End Training
  11. Experiments
  12. Datasets and Evaluation Protocol
  13. Implementation Details
  14. Results
  15. Runtime Analysis
  16. ...and 2 more sections

Figures (2)

  • Figure 1: HOTFLoc++ achieves Pareto-optimality for place recognition (top) and metric localisation (bottom) on CS-Wild-Places. Filled symbols denote results after re-ranking.
  • Figure 2: Pipeline of HOTFLoc++. (a) We use an octree-based transformer backbone griffithsHOTFormerLocHierarchicalOctree2025 to extract multi-scale local features and a robust global descriptor for place recognition. (b) Our learnable Multi-Scale Geometric Verification module re-ranks retrievals, improving robustness to degraded or erroneous single-resolution correspondences. (c) Our coarse-to-fine registration extracts patch-level correspondences and refines the patch-wise registration which maximises the global inlier ratio.