Research map / Point clouds / Point cloud classification
ForestLPR: LiDAR Place Recognition in Forests Attentioning Multiple BEV Density Images
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 ForestLPR, a LiDAR place recognition method for forests that feeds multiple height-sliced BEV density images through a DeiT-S backbone and fuses them with a patch-level multi-BEV interaction module.
Problem
LiDAR place recognition in natural forests, where high self-similarity and seasonal/growth-driven vegetation changes make recognition hard and urban methods generalize poorly.
Approach
Represent each scan as a set of BEV density images from horizontal slices at different heights, process them with a shared transformer, and adaptively weight patch-level features across heights via a multi-BEV interaction module, after ground removal and height normalization that discards seasonal canopy and ground variation.
Key results
- Achieves an average 7.38% Recall@1 gain over the closest competitor on intra-sequence loop closure detection across all datasets.
- Achieves an average 9.11% Recall@1 gain on inter-sequence re-localization on Wild-Places over the top-2 baselines.
- Ablation: replacing the multi-BEV interaction module with concatenation reduces average R@1 by 13.96%, and removing interaction reduces it by 6.10%.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use multiple BEV density images rather than a single BEV image (ablation) | ForestLPR · Wild-Places V-03 | Recall@1: 53.70 -> 76.53 | Helped |
| use multiple BEV density images rather than a single BEV image (ablation) | ForestLPR · Wild-Places V-04 | Recall@1: 64.46 -> 82.33 | Helped |
| use multiple BEV density images rather than a single BEV image (ablation) | ForestLPR · Wild-Places K-03 | Recall@1: 54.72 -> 74.89 | Helped |
| use multiple BEV density images rather than a single BEV image (ablation) | ForestLPR · Wild-Places K-04 | Recall@1: 70.03 -> 76.73 | Helped |
| use multiple BEV density images rather than a single BEV image (ablation) | ForestLPR · ANYmal | Recall@1: 58.51 -> 71.87 | Helped |
| extract features from concatenated BEV images instead of separate images with interaction (ablation) | ForestLPR · Wild-Places V-03 | Recall@1: 68.17 -> 76.53 | Helped |
| extract features from concatenated BEV images instead of separate images with interaction (ablation) | ForestLPR · Wild-Places V-04 | Recall@1: 64.25 -> 82.33 | Helped |
| extract features from concatenated BEV images instead of separate images with interaction (ablation) | ForestLPR · Wild-Places K-03 | Recall@1: 55.84 -> 74.89 | Helped |
| extract features from concatenated BEV images instead of separate images with interaction (ablation) | ForestLPR · Wild-Places K-04 | Recall@1: 71.04 -> 76.73 | Helped |
| extract features from concatenated BEV images instead of separate images with interaction (ablation) | ForestLPR · ANYmal | Recall@1: 63.26 -> 71.87 | Helped |
| add interaction-based weighting instead of max pooling across height descriptors (ablation) | ForestLPR · Wild-Places V-03 | Recall@1: 59.68 -> 76.53 | Helped |
| add interaction-based weighting instead of max pooling across height descriptors (ablation) | ForestLPR · Wild-Places V-04 | Recall@1: 76.87 -> 82.33 | Helped |
Surprising findings
- Concatenating multiple BEV images at input performs worse than separate per-slice feature extraction, indicating feature-level interaction in the backbone is insufficient.
Limitations
- Only tested on forest-like datasets; generalization to dense jungles is uncertain.
- Some canopy and bushes remain in the cropped point cloud, and blind spots in datasets require custom overlap-based positive mining.
Experimental setup
Datasets: Wild-Places ANYmal Dataset Botanic Dataset
Benchmarks: Wild-Places ANYmal Dataset Botanic Dataset
Metrics: Recall@1 maximum F1 score latency memory footprint
Baselines: PointPillar TransLoc3D MinkLoc3Dv2 LoGG3D-Net Scan-Context BEVPlace MapClosures
Builds on: DeiT GeM pooling Scan Context BEVPlace
Closest papers by meaning
- HOTFormerLoc: Hierarchical Octree Transformer for Versatile Lidar Place Recognition Across Ground and Aerial Views (CVPR 2025 · similarity 0.89)
- L2RSI: Cross-view LiDAR-based Place Recognition for Large-scale Urban Scenes via Remote Sensing Imagery (NeurIPS 2025 · similarity 0.87)
- LightLoc: Learning Outdoor LiDAR Localization at Light Speed (CVPR 2025 · similarity 0.87)
- LiSA: LiDAR Localization with Semantic Awareness (CVPR 2024 · similarity 0.85)
- FoMo: Multi-Modal, Multi-Scale and Multi-Task Remote Sensing Foundation Models for Forest Monitoring (AAAI 2025 · similarity 0.84)
- DUNIA: Pixel-Sized Embeddings via Cross-Modal Alignment for Earth Observation Applications (ICML 2025 · similarity 0.84)
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.