Research map / Vision-language and multimodal / Cross-modal learning
MMLandmarks: a Cross-View Instance-Level Benchmark for Geo-Spatial Understanding
Oskar Kristoffersen, Alba Reinders Sánchez, Morten Rieger Hannemose, Anders Bjorholm Dahl, Dim P. Papadopoulos
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| sample latest aerial image and train on outdoor-only ground subset (ablation) | MMCLIP · MMLandmarks Satellite-to-GPS | distance accuracy at 1 km (%): 27.2 -> 36.9 | Improved |
| use latest instead of random satellite image, with outdoor subset (ablation) | MMCLIP · MMLandmarks Satellite-to-Ground | mAP@1k: 17.45 -> 18.79 | Improved |
| increase number of modalities (add text and GPS) (ablation) | MMCLIP · MMLandmarks Ground-to-Satellite | mAP@1k: 25.59 -> 24.94 | No clear effect |
| use first sentence vs random sentence for text sampling (ablation) | MMCLIP · MMLandmarks Satellite-to-Ground | mAP@1k: 17.39 vs 16.93 | No clear effect |
| use all-pairwise contrastive objective instead of ground-anchored ImageBind-style objective (ablation) | MMCLIP · MMLandmarks Satellite-to-GPS | distance accuracy at 1 km (%): 18.3 -> 36.9 | Improved |
| use ground-anchored objective instead of all pairwise (ablation) | MMCLIP · MMLandmarks retrieval | mAP@1k ground-to-satellite: 26.20 -> 27.46 | Improved |
| filter indoor ground images from training (ablation) | MMCLIP · MMLandmarks Satellite-to-Ground | mAP@1k: 16.93 -> 17.45 | Improved |
| use specialized cross-view models zero-shot | Uni-1652, TransGeo, Sample4Geo · MMLandmarks Ground-to-Satellite | R@1: 0.2-0.85 vs 14.6 for SigLIP2 | Hurt |
| scale foundation model from ViT-B to ViT-L and raise resolution | SigLIP2 zero-shot · MMLandmarks Ground-to-Satellite | R@1: 8.5 -> 14.6 | Improved |
| train baseline on MMLandmarks versus zero-shot CLIP-like models | CLIP ViT-L/336 · MMLandmarks Ground-to-Satellite | R@1: 11.9 -> 20.5 | Improved |
Closest papers by meaning
- G3: An Effective and Adaptive Framework for Worldwide Geolocalization Using Large Multi-Modality Models (NeurIPS 2024 · similarity 0.88)
- GeoRanker: Distance-Aware Ranking for Worldwide Image Geolocalization (NeurIPS 2025 · similarity 0.87)
- GeoViS: Geospatially Rewarded Visual Search for Remote Sensing Visual Grounding (CVPR 2026 · similarity 0.87)
- TimeSpot: Benchmarking Geo-Temporal Understanding in Vision–Language Models in Real-World Settings (ICML 2026 · similarity 0.87)
- GeoRouter: Dynamic Paradigm Routing for Worldwide Image Geolocalization (NeurIPS 2026 · similarity 0.87)
- OpenStreetView-5M: The Many Roads to Global Visual Geolocation (CVPR 2024 · similarity 0.87)
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.
