Geo$^\textbf{2}$: Geometry-Guided Cross-view Geo-Localization and Image Synthesis

Abstract: Cross-view geo-spatial learning consists of two important tasks: Cross-View Geo-Localization (CVGL) and Cross-View Image Synthesis (CVIS), both of which rely on establishing geometric correspondences between ground and aerial views. Recent Geometric Foundation Models (GFMs) have demonstrated strong capabilities in extracting generalizable 3D geometric features from images, but their potential in cross-view geo-spatial tasks remains underexplored. In this work, we present Geo^2, a unified framework that leverages Geometric priors from GFMs (e.g., VGGT) to jointly perform geo-spatial tasks, CVGL and bidirectional CVIS. Despite the 3D reconstruction ability of GFMs, directly applying them to CVGL and CVIS remains challenging due to the large viewpoint gap between ground and aerial imagery. We propose GeoMap, which embeds ground and aerial features into a shared 3D-aware latent space, effectively reducing cross-view discrepancies for localization. This shared latent space naturally bridges cross-view image synthesis in both directions. To exploit this, we propose GeoFlow, a flow-matching model conditioned on geometry-aware latent embeddings. We further introduce a consistency loss to enforce latent alignment between the two synthesis directions, ensuring bidirectional coherence. Extensive experiments on standard benchmarks, including CVUSA, CVACT, and VIGOR, demonstrate that Geo^2 achieves state-of-the-art performance in both localization and synthesis, highlighting the effectiveness of 3D geometric priors for cross-view geo-spatial learning.

• The paper presents Geo2, a unified framework leveraging geometric priors to jointly tackle cross-view geo-localization and bidirectional image synthesis.
• GeoMap uses dual-branch networks with 3D-aware feature extraction and cross-attention mechanisms to improve retrieval under varied viewpoints.
• GeoFlow applies flow matching conditioned on latent embeddings to enable efficient, high-quality image synthesis validated by benchmark metrics.

"Geo$^\text{2}$: Geometry-Guided Cross-view Geo-Localization and Image Synthesis" (2603.25819)

Introduction

Cross-view geo-spatial learning encompasses two pivotal tasks: Cross-View Geo-Localization (CVGL) and Cross-View Image Synthesis (CVIS). Both tasks rely heavily on establishing geometric correspondences between ground-level and aerial views. CVGL aims at identifying the location of a query ground image by matching it against a database of geo-tagged aerial images. Conversely, CVIS involves synthesizing either a satellite view from a ground view (denoted as G2S) or a ground view from a satellite view (S2G). The challenge lies in bridging the substantial viewpoint and geometric disparities between these images. Previous methods introduce geometric cues and customized modules to capture structural similarities yet fail to unify CVGL and CVIS within a single framework.

Geometric Foundation Models (GFMs) like VGGT have shown promising abilities in extracting 3D features across diverse visual tasks but remain largely unexplored in cross-view geo-spatial learning. This paper proposes Geo$^\text{2}$, a unified framework leveraging geometric priors to jointly address CVGL and bidirectional CVIS, effectively embedding images into a shared geometry-aware latent space for enhanced embedding alignment and consistency.

Figure 1: Illustration of directly using VGGT on satellite (a) and ground (c) images, leading to incorrect reconstructed results shown in (b).

Methodology

Geo$^\text{2}$ Overview

The Geo$^\text{2}$ framework impacts both CVGL and CVIS tasks by embedding images into a shared latent space using two modules: GeoMap and GeoFlow. GeoMap utilizes dual-branch models to encode ground and aerial images into geometry-aware features. Subsequently, these embeddings streamline retrieval processes between spatial domains, enhancing CVGL reliability.

GeoFlow, a flow-matching model conditioned on latent embeddings, facilitates bidirectional image synthesis without requiring dedicated retraining for each synthesis direction. By applying a consistency loss during training, the model ensures coherent geometry between synthesis directions and geometric alignment between views.

Figure 2: Overview of the Geo$^\text{2}$ framework, illustrating the extraction of geometric features embedded into a shared latent space and their utilization for CVGL and CVIS.

Geo-localization using GeoMap

GeoMap enhances CVGL by aggregating high-dimensional features and utilizing cross-attention to align embeddings more intimately. By leveraging VGGT's dense 3D-aware features and enriching them with semantic information from complementary networks like ConvNeXt, GeoMap ensures robust retrieval even under substantial geometric variance.

Cross-view Image Synthesis with GeoFlow

GeoFlow reformulates the conventional image synthesis problem as domain translation, incorporating flow matching to achieve efficient bidirectional synthesis. By integrating geometry-aware features from GeoMap, it produces high-quality image pairs in both synthesis directions, validated through significant improvements in perceptual and structural metrics compared to baseline models.

Figure 3: GeoFlow pipeline illustrating the integration of geometry-aware features in bidirectional image synthesis conditioned on the shared latent space.

Experiments and Results

Extensive experiments on benchmark datasets (CVUSA, CVACT, VIGOR) show that Geo$^\text{2}$ achieves state-of-the-art performance on both CVGL and CVIS. The framework demonstrates superior retrieval accuracy and remarkable synthesis quality, benefiting from the incorporated geometric priors that enhance the robustness of feature representations against viewpoint variability.

Figure 4: Visualization of generated images from Geo$^\text{2}$ on the CVUSA dataset. From left to right: ground truth satellite image, ground truth ground image, Ground-to-Satellite generated image, Satellite-to-Ground generated image.

Conclusion

Geo$^\text{2}$ introduces a novel approach to cross-view geo-spatial learning by synergistically employing geometric foundations within a unified framework. The joint exploration of CVGL and CVIS not only advances individual task performance but also indicates future research directions leveraging geometric priors for multi-task spatial reasoning and synthesis. The implications of Geo$^\text{2}$ are profound, laying groundwork for more integrated geo-spatial perception systems and enhanced AI-driven applications in localization and image synthesis.