Research map / Vision-language and multimodal / Cross-modal learning
Multi-Modal Aerial-Ground Cross-View Place Recognition with Neural ODEs
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 AGPlace, a multi-modal (image + LiDAR) ground place recognition model that matches aerial maps using a neural-ODE fusion embedding on a manifold and a multi-view multi-modal alignment loss.
Problem
Aerial-ground cross-view place recognition where ground multi-modal queries (images + point clouds) must be matched against aerial database maps, a setting that prior multi-modal PR work (ground-to-ground) and aerial-ground PR work (single-modal images) have not addressed.
Approach
A third manifold branch builds a fusion embedding by evolving states through neural ODEs across backbone blocks (L to 1), with manifold chart functions linking 2D/3D features to the fusion manifold; the embedding then guides modal-wise decoding, trained with a multi-view multi-modal (MVMM) binary alignment loss plus triplet loss.
Key results
- On KITTI360-AG with satellite database, AGPlace reaches R@1 32.0 / R@5 47.6 / R@10 54.9, above the best baseline ArcGeo (28.2 R@1).
- On nuScenes-AG under LiDAR failure, AGPlace achieves R@1 12.9 vs 8.6 for LCPR.
- On Oxford RobotCar, AGPlace achieves 99.9 AR@1% and 99.3 AR@1 on the Oxford-Ada split, exceeding prior ground-ground SOTA.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use AGPlace instead of the strongest listed baseline | Aerial-ground PR · KITTI360-AG (satellite) | R@1: 29.9 -> 32.0 | Helped |
| Use AGPlace instead of the strongest listed baseline | Aerial-ground PR · KITTI360-AG (road map) | R@1: 27.2 -> 28.2 | Helped |
| Use AGPlace instead of the strongest listed baseline | Aerial-ground PR · nuScenes-AG (cams + LiDAR) | R@1: 71.9 -> 75.6 | Helped |
| Add fusion embedding construction (stage 1) (ablation) | AGPlace · KITTI360-AG (satellite) | R@1: 28.5 -> 32.0 | Helped |
| Add modal-wise fusion (stage 2) (ablation) | AGPlace · KITTI360-AG (satellite) | R@1: 30.6 -> 32.0 | Helped |
| Add the MVMM loss (ablation) | AGPlace · KITTI360-AG (satellite) | R@1: 30.9 -> 32.0 | Helped |
| Use ODE state updating instead of MLP (ablation) | AGPlace · KITTI360-AG (satellite) | R@1: 30.0 -> 32.0 | Helped |
| Use ODE state updating instead of attention (ablation) | AGPlace · KITTI360-AG (satellite) | R@1: 30.4 -> 32.0 | Helped |
| Reverse fusion direction from block L → block 1 to block 1 → block L (ablation) | AGPlace · KITTI360-AG (satellite) | R@1: 32.0 -> 31.7 | Hurt |
| Add road map aerial input alongside satellite imagery (ablation) | AGPlace · KITTI360-AG | R@1: 32.0 -> 34.7 | Helped |
| Add modal-wise fusion (stage 2) (ablation) | AGPlace · nuScenes-AG (cams fail) | R@1: 19.2 -> 22.8 | Helped |
Surprising findings
- Road-map aerial databases already give considerable recall, with most models performing comparably to satellite databases.
- Directly adapting ground-ground PR baselines to the cross-view setting performs worse than AGPlace.
Limitations
- Focuses on ground query data collected in urban environments, which may limit generalizability.
- Runtime is 62 FPS, slower than several multi-modal baselines (e.g., 80 FPS MinkLoc++).
Experimental setup
Datasets: KITTI360 nuScenes Oxford RobotCar Google Maps Static API aerial imagery
Benchmarks: KITTI360-AG nuScenes-AG Oxford RobotCar
Metrics: R@1 R@5 R@10 AR@1% AR@1
Baselines: MinkLoc++ AdaFusion MSSPlace LCPR UMF Sample4Geo ArcGeo TransGeo SALAD AnyLoc Lip-Loc VXP
Builds on: MinkLoc++ Sample4Geo TransGeo NetVLAD triplet loss
Closest papers by meaning
- L2RSI: Cross-view LiDAR-based Place Recognition for Large-scale Urban Scenes via Remote Sensing Imagery (NeurIPS 2025 · similarity 0.85)
- DistilVPR: Cross-Modal Knowledge Distillation for Visual Place Recognition (AAAI 2024 · similarity 0.85)
- View From Above: Orthogonal-View aware Cross-view Localization (CVPR 2024 · similarity 0.85)
- Weather-Robust Cross-View Geo-Localization via Prototype-Based Semantic Part Discovery (NeurIPS 2026 · similarity 0.85)
- From Coarse to Fine: A Matching and Alignment Framework for Unsupervised Cross-View Geo-Localization (AAAI 2025 · similarity 0.84)
- BevSplat: Resolving Height Ambiguity via Feature-Based Gaussian Primitives for Weakly-Supervised Cross-View Localization (NeurIPS 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.