Research map / 3D vision / Depth and stereo

Seeing Across Skies and Streets: Feedforward 3D Reconstruction from Satellite, Drone, and Ground Images

NeurIPS 2026 · 3D vision › Depth and stereo · approach: Foundation models for monocular geometry · 0 citations (OpenAlex)

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.

Cross3R is a feed-forward model that jointly reconstructs satellite, UAV and ground views (point clouds, 6-DoF poses) and localizes ground/UAV cameras on the satellite tile, trained on a new CrossGeo tri-view dataset.

Problem

Cross-view localization is usually limited to 3-DoF (x, y, yaw) because a nadir satellite tile gives no cues for roll, pitch or altitude, and existing feed-forward 3D models fail to align satellite, UAV and ground point clouds.

Approach

Adds a UAV image as an intermediate viewpoint and extends pi^3 with a satellite-specific orthographic head predicting a per-tile learned relative scale rho, Fourier positional embedding for satellite tokens, satellite-to-perspective feature injection, and per-sample altitude redefinition, enabling 6-DoF cross-view localization without satellite intrinsics.

Key results

• On CrossGeo ground camera, Cross3R reduces mean meter error to 3.68 m versus 7.16 m for pi^3* and raises ground PCK@2m to 42.60% from 11.04%.
• Zero-shot on KITTI Test 2 (no orientation prior), Cross3R achieves 11.69 m mean localization error and 85.95% lateral recall at 5 m, outperforming CCVPE, which is trained on KITTI.
• On AnyVisLoc (OOD real UAV), Cross3R has the best ground meter-mean (10.49 m) and UAV meter-mean (14.51 m) among compared methods.

Surprising findings

• Zero-shot Cross3R trained only on CrossGeo beats dedicated KITTI-trained cross-view methods on most KITTI metrics.
• Training with the raw satellite altitude (~5,726 m) was unstable and diverged in the authors' runs.

Limitations

• At most three images per forward pass due to limited compute.
• Handles only pinhole cameras; no panoramic or fisheye support.
• Data-hungry; performance still improving with more data, and AnyVisLoc results are noisier qualitatively due to out-of-distribution UAV altitudes above 120 m.

Experimental setup

Datasets: CrossGeo AnyVisLoc KITTI

Benchmarks: CrossGeo KITTI Test 2 AnyVisLoc

Metrics: Acc-mean delta threshold accuracy RRA RTA AUC@30 Meter mean/median Yaw mean/median PCK@2m PCK@5m lateral/longitudinal recall

Baselines: AerialMD VGGT pi^3 LM SliceMatch CCVPE

Builds on: pi^3 DUSt3R Depth Anything v3 Prior Depth Anything CCVPE

Closest papers by meaning

• Weather-Robust Cross-View Geo-Localization via Prototype-Based Semantic Part Discovery (NeurIPS 2026 · similarity 0.89)
• View From Above: Orthogonal-View aware Cross-view Localization (CVPR 2024 · similarity 0.87)
• BevSplat: Resolving Height Ambiguity via Feature-Based Gaussian Primitives for Weakly-Supervised Cross-View Localization (NeurIPS 2025 · similarity 0.87)
• FG^2: Fine-Grained Cross-View Localization by Fine-Grained Feature Matching (CVPR 2025 · similarity 0.87)
• Game4Loc: A UAV Geo-Localization Benchmark from Game Data (AAAI 2025 · similarity 0.86)
• AerialMegaDepth: Learning Aerial-Ground Reconstruction and View Synthesis (CVPR 2025 · similarity 0.86)