Research map / Point clouds / Point cloud classification
HOTFormerLoc: Hierarchical Octree Transformer for Versatile Lidar Place Recognition Across Ground and Aerial Views
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes HOTFormerLoc, a hierarchical octree transformer with cylindrical octree windows, relay tokens and pyramid attentional pooling for ground-to-ground and ground-to-aerial lidar place recognition, and introduces the CS-Wild-Places cross-source forest dataset.
Problem
Lidar place recognition across ground and aerial viewpoints in urban and forest environments, where point density varies with distance from a spinning lidar and cross-source point distributions differ.
Approach
Combines octree-serialised window attention with relay tokens that give multi-scale global-local interaction, cylindrical (rather than Cartesian) octree partitioning that matches spinning-lidar density, a distribution-aware absolute positional encoding, and pyramid attentional pooling into a global descriptor.
Key results
- On CS-Wild-Places, HOTFormerLoc improves AR@1 by 5.5%-11.5% on the Baseline and Unseen sets over SOTA.
- On CS-Campus3D, HOTFormerLoc reaches 80.4 AR@1 vs 70.7 for CrossLoc3D.
- On Oxford RobotCar, HOTFormerLoc reaches 92.1 mean AR@1 vs 90.0 for MinkLoc3Dv2 (96.4 on Oxford subset).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use HOTFormerLoc instead of CrossLoc3D | CS-Campus3D ground-query, aerial-database retrieval · CS-Campus3D | AR@1: 70.7 -> 80.4 | Helped |
| Use HOTFormerLoc instead of CrossLoc3D | CS-Campus3D ground-query, aerial-database retrieval · CS-Campus3D | AR@1%: 85.7 -> 94.9 | Helped |
| Use HOTFormerLoc instead of LoGG3D-Net | Wild-Places · Wild-Places mean | AR@1: 77.3 -> 74.8 | Hurt |
| Use HOTFormerLoc instead of MinkLoc3Dv2 | Oxford RobotCar · Oxford RobotCar mean | AR@1: 90.0 -> 92.1 | Helped |
| Disable relay tokens (ablation) | HOTFormerLoc · Oxford | AR@1 (Mean): -2.5% | Hurt |
| Disable relay tokens (ablation) | HOTFormerLoc · CS-Campus3D | AR@1: -4.5% | Hurt |
| Disable relay tokens (ablation) | HOTFormerLoc · CS-Wild-Places | AR@1 (Mean): -4.7% | Hurt |
| Disable ADaPE (ablation) | HOTFormerLoc · CS-Campus3D | AR@1: -2.9% | Hurt |
| Use L=2 pyramid levels (ablation) | HOTFormerLoc · CS-Campus3D | AR@1: -5.1% | Hurt |
| Replace pyramid attentional pooling with GeM pooling (ablation) | HOTFormerLoc · CS-Campus3D | AR@1: -22.5% | Hurt |
| Replace pyramid attentional pooling with Pyramid GeM pooling (ablation) | HOTFormerLoc · CS-Wild-Places | AR@1 (Mean): -12.3% | Hurt |
| Use octree depth 7 and window size 64 instead of depth 6 and window size 64 (ablation) | HOTFormerLoc · CS-Wild-Places | AR@1 (Mean): 34.5 -> 58.3 | Helped |
Surprising findings
- LoGG3D-Net, the top performer on Wild-Places, drops considerably on CS-Wild-Places, which the authors attribute to its local-consistency loss being ill-suited to cross-source data.
- Increasing octree depth beyond 7 does not improve performance, attributed to redundancy in deeper octrees.
Limitations
- Bulkier and slower than sparse-CNN baselines (35.4M params, 270 ms inference).
- CrossLoc3D requires fixed 4096-point inputs, so it had to be retrained on downsampled data for comparison.
- Cross-source LPR remains challenging; the authors note token pruning as a needed improvement.
Experimental setup
Datasets: Oxford RobotCar CS-Campus3D Wild-Places CS-Wild-Places
Benchmarks: CS-Wild-Places CS-Campus3D Wild-Places Oxford RobotCar
Metrics: AR@1 AR@1% MRR recall@N parameter count inference time
Baselines: PointNetVLAD PPT-Net TransLoc3D MinkLoc3Dv2 CrossLoc3D LoGG3D-Net GAPR
Builds on: OctFormer MinkLoc3Dv2 Swin Transformer Focal Transformer FasterViT
Closest papers by meaning
- ForestLPR: LiDAR Place Recognition in Forests Attentioning Multiple BEV Density Images (CVPR 2025 · similarity 0.89)
- L2RSI: Cross-view LiDAR-based Place Recognition for Large-scale Urban Scenes via Remote Sensing Imagery (NeurIPS 2025 · similarity 0.88)
- UniMamba: Unified Spatial-Channel Representation Learning with Group-Efficient Mamba for LiDAR-based 3D Object Detection (CVPR 2025 · similarity 0.86)
- LightLoc: Learning Outdoor LiDAR Localization at Light Speed (CVPR 2025 · similarity 0.86)
- LiDAR-Net: A Real-scanned 3D Point Cloud Dataset for Indoor Scenes (CVPR 2024 · similarity 0.86)
- GBlobs: Explicit Local Structure via Gaussian Blobs for Improved Cross-Domain LiDAR-based 3D Object Detection (CVPR 2025 · similarity 0.85)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.