Aerial, Satellite & BEV Perception
Robust Cross-View Perception and Localization Without Paired Ground-Level Street Imagery
Generated automatically from the limitations stated in 4 papers (ECCV, EMNLP, CVPR), listed under Evidence. It is not a paper, and it does not come from papers submitted to CSPaper.
The problem
Current cross-view visual geolocalization, 3D scene reconstruction, and BEV perception models strictly depend on dense, spatially calibrated ground-level panoramas paired with satellite imagery or pre-existing 3D city meshes. In rural, sparsely populated, or underdeveloped regions, ground-level street-view coverage is either completely absent or uncalibrated. This reliance on dense ground-level imagery structurally restricts cross-view perception algorithms to major metropolitan areas with dedicated mapping fleets.
Why it matters
Enables visual localization, BEV understanding, and 3D urban mapping in unmapped rural regions, developing nations, and disaster-affected zones where ground-level panoramic survey infrastructure does not exist.
Ways to approach it
Prior-work checks are free with an account. Results someone already ran are shown to everyone.
- 1
Missing-modality robustness benchmark: Systematically evaluate existing cross-view retrieval, BEV mapping, and visual geolocalization models under controlled degradations of ground-view density, missing orientation calibration, and complete absence of street-view data. Measure the decay of Top-1/Top-5 retrieval recall and BEV layout IoU as paired ground data is progressively removed across diverse geographic regions.
- 2
Unpaired and weakly-supervised cross-view alignment: Train cross-view feature representations using overhead satellite/aerial imagery matched against sparse, uncalibrated crowdsourced photos (e.g., Mapillary / OpenStreetMap) without requiring rigid 1-to-1 spatial calibration or high-overlap sky constraints. Measure geolocalization recall and cross-view retrieval accuracy against fully calibrated baseline models.
- 3
Overhead-only ground geometry and facade estimation: Evaluate learning-based models that predict ground-level appearance, road layouts, and facade structures directly from multi-angle satellite imagery without relying on city-specific ground priors or pre-existing 3D meshes. Measure structural Chamfer distance and facade semantic accuracy against full cross-view baseline methods.
Have a different approach?
Describe how you would tackle this problem and we'll look for papers that already do it. Free; your text stays private.
Why it might fail
If overhead-only satellite resolution is fundamentally insufficient to resolve ground-level ambiguities (e.g., under building overhangs or dense tree canopies), or if large foundation vision models trained on Internet-scale uncurated web images make explicit cross-view spatial alignment trivial without specialized research architectures.
Evidence
Each paper's own statement of the limitation, verbatim.
- AI Knows Where You Are: Exposure, Bias, and Inference in Multimodal Geolocation with KoreaGEOEMNLP 2025
Relies on publicly available street-view imagery, so extremely underdeveloped or sparsely populated areas lack usable images and are underrepresented
- SG-BEV: Satellite-Guided BEV Fusion for Cross-View Semantic SegmentationCVPR 2024
Requires paired, spatially calibrated street-view panoramas and satellite images with known correspondence (and offsets, as in Vigor); not applicable where such pairs are unavailable
- MeshVPR: Citywide Visual Place Recognition Using 3D MeshesECCV 2024
Meshes built from aerial imagery lack accurate facades/vegetation detail, and coverage is limited to cities with freely available 3D meshes
- Geospecific View Generation - Geometry-Context Aware High-resolution Ground View Inference from Satellite ViewsECCV 2024
Requires multi-view satellite imagery with stereo matching, accurate geolocation of street-view pairs (filtered to >95% sky-overlap ratio), and region-specific street-view data (~500 images per city) to train each geospecific prior
Nearest existing work
- Cross-view image geo-localization with Panorama-BEV Co-Retrieval NetworkECCV 2024
- BevSplat: Resolving Height Ambiguity via Feature-Based Gaussian Primitives for Weakly-Supervised Cross-View LocalizationNeurIPS 2025
- Wide-Area Image Geolocalization With Aerial Reference ImageryICCV 2015
- SG-BEV: Satellite-Guided BEV Fusion for Cross-View Semantic SegmentationCVPR 2024
- Uncertainty-Aware Vision-Based Metric Cross-View GeolocalizationCVPR 2023
- Coming Down to Earth: Satellite-to-Street View Synthesis for Geo-LocalizationCVPR 2021
- Geospecific View Generation - Geometry-Context Aware High-resolution Ground View Inference from Satellite ViewsECCV 2024
- View From Above: Orthogonal-View aware Cross-view LocalizationCVPR 2024
- Spatial-Aware Feature Aggregation for Image based Cross-View Geo-LocalizationNeurIPS 2019
- VIGOR: Cross-View Image Geo-Localization Beyond One-to-One RetrievalCVPR 2021
- Statewide Visual Geolocalization in the WildECCV 2024
- Leveraging BEV Paradigm for Ground-to-Aerial Image SynthesisICCV 2025
- GeoBridge: A Semantic-Anchored Multi-View Foundation Model Bridging Images and Text for Geo-LocalizationCVPR 2026
- Beyond Cross-View Image Retrieval: Highly Accurate Vehicle Localization Using Satellite ImageCVPR 2022
- Boosting 3-DoF Ground-to-Satellite Camera Localization Accuracy via Geometry-Guided Cross-View TransformerICCV 2023
Related open problems
Aerial, Satellite & BEV Perception
Benchmarking and Mitigating UAV Perception Failure Under Rapid Motion Blur, Scale Shifts, and Occlusion
Current perception models deployed on unmanned aerial vehicles achieve low absolute accuracy on standard downstream tasks (e.g., recall under 7.5% in scene understanding, under 40% in action recognition, and under 30% m_vIoU in tracking). These failures are driven by real-world flight conditions—rapid camera motion blur, extreme resolution and altitude shifts, severe occlusion, and small target sizes—that standard benchmarks do not systematically couple. As a consequence, aerial vision pipelines remain brittle and cannot be reliably trusted for autonomous navigation, search-and-rescue, or monitoring in complex flight regimes.
Aerial, Satellite & BEV Perception
Robustness and Degradation Benchmarking of Map-Conditioned Perception Under Imperfect OpenStreetMap Priors
Contemporary BEV perception, aerial localization, and landmark-guided navigation models increasingly condition their visual representations on OpenStreetMap (OSM) vector geometry and metadata. In practice, OSM data exhibits severe geographic disparities, missing road classes, topological errors, and incomplete landmark tags. Because existing methods assume clean and complete vector maps as an operational precondition, system performance degrades unpredictably in under-mapped or rural areas, preventing deployment outside densely curated metropolitan regions.