TL;DR
This paper proposes a single-stage framework and a new large-scale dataset to accurately locate objects in satellite images from ground or drone views, significantly outperforming existing 2D matching methods.
Problem
Existing cross-view object geo-localization (CVOGL) methods heavily rely on 2D appearance matching and are constrained by limited datasets lacking geometric metadata, diverse prompts, and standard field-of-view imagery.
Approach
New Dataset (GeoSet): Introduces a large-scale, high-fidelity building dataset with over 220,000 ground-satellite and drone-satellite pairs, providing multi-modal prompts and camera poses for flexible target referring and explicit spatial modeling.
Single-Stage Framework (GAGeo): Proposes a novel framework built upon a permutation-equivariant 3D foundation model ($π^3$). It integrates visual features, referring prompts, and learnable task tokens to jointly predict bounding boxes, segmentation masks, and camera poses in a single forward pass.
Contrastive Loss: Introduces a contrastive loss that uses the satellite view as a universal anchor, implicitly aligning ground and drone representations to enable zero-shot ground-to-drone localization without requiring triplet training data.
Results & Contribution
Performance: Extensive experiments show the approach significantly outperforms state-of-the-art methods.
Generalization: Demonstrates exceptional generalization ability in unseen scenes and novel cross-view setups.