Satellite-Free Training for Drone-View Geo-Localization

Abstract: Drone-view geo-localization (DVGL) aims to determine the location of drones in GPS-denied environments by retrieving the corresponding geotagged satellite tile from a reference gallery given UAV observations of a location. In many existing formulations, these observations are represented by a single oblique UAV image. In contrast, our satellite-free setting is designed for multi-view UAV sequences, which are used to construct a geometry-normalized UAV-side location representation before cross-view retrieval. Existing approaches rely on satellite imagery during training, either through paired supervision or unsupervised alignment, which limits practical deployment when satellite data are unavailable or restricted. In this paper, we propose a satellite-free training (SFT) framework that converts drone imagery into cross-view compatible representations through three main stages: drone-side 3D scene reconstruction, geometry-based pseudo-orthophoto generation, and satellite-free feature aggregation for retrieval. Specifically, we first reconstruct dense 3D scenes from multi-view drone images using 3D Gaussian splatting and project the reconstructed geometry into pseudo-orthophotos via PCA-guided orthographic projection. This rendering stage operates directly on reconstructed scene geometry without requiring camera parameters at rendering time. Next, we refine these orthophotos with lightweight geometry-guided inpainting to obtain texture-complete drone-side views. Finally, we extract DINOv3 patch features from the generated orthophotos, learn a Fisher vector aggregation model solely from drone data, and reuse it at test time to encode satellite tiles for cross-view retrieval. Experimental results on University-1652 and SUES-200 show that our SFT framework substantially outperforms satellite-free generalization baselines and narrows the gap to methods trained with satellite imagery.

• The paper presents a Satellite-Free Training framework that leverages 3D Gaussian Splatting to reconstruct dense scenes from drone images without satellite data.
• The paper employs geometry-guided inpainting and Fisher vector aggregation with a drone-trained GMM to enable robust cross-view retrieval, yielding up to 93.50% R@1.
• The paper validates the approach as a viable solution for GPS-denied environments by substantially narrowing the performance gap compared to satellite-trained methods.

Satellite-Free Training for Drone-View Geo-Localization

Problem Definition and Motivation

The paper "Satellite-Free Training for Drone-View Geo-Localization" (2604.01581) addresses the challenge of matching UAV (drone) visual observations to satellite imagery—a critical capability for localization in GPS-denied environments. Conventional Drone-View Geo-Localization (DVGL) approaches rely on satellite imagery during training, either through paired supervision or unsupervised alignment. However, satellite data may not always be available due to confidentiality, transmission bottlenecks, or restricted access. This paper tackles the underexplored setting where DVGL models are trained without any satellite images, requiring the model to generalize from drone views alone and to accurately retrieve satellite tiles at test time.

Proposed Satellite-Free Training Framework

The core contribution is a Satellite-Free Training (SFT) framework composed of a multi-stage pipeline that transforms multi-view drone sequences into cross-view compatible representations, without any satellite data in the representation learning pipeline. The framework comprises several key components:

1. 3D Scene Reconstruction with 3D Gaussian Splatting: Multi-view UAV images are used to reconstruct a dense, view-consistent 3D scene, leveraging advances in 3D Gaussian Splatting (3DGS) for high-fidelity geometry and appearance modeling, initialized with structure-from-motion (COLMAP) outputs.
2. Pseudo-Orthophoto Generation: The reconstructed Gaussian field is rendered into a pseudo-orthophoto from a canonical top-down viewpoint using PCA-guided ground plane estimation and orthographic projection. A soft-roof compositing strategy integrates multiple depth layers, mitigating perspective distortions and preserving structural cues.
3. Geometry-Guided Inpainting: To address texture incompleteness resulting from occlusions and reconstruction gaps, the pipeline applies a hybrid inpainting approach. Morphology-based detection and classical KNN/Telea inpainting fill small holes, while the LaMa deep completion model fills larger gaps—processing is geometry aligned to respect scene semantics.
4. Satellite-Free Representation Learning: A frozen, web-pretrained DINOv3 Vision Transformer extracts patch descriptors from the completed pseudo-orthophotos. A Gaussian Mixture Model (GMM) is trained exclusively on these descriptors to build a drone-only visual vocabulary. Image-level representations are then obtained via Fisher vector aggregation, encoding higher-order statistics of the patch distribution.
5. Cross-View Retrieval: At inference, satellite tiles (now available) are encoded with the same DINOv3 backbone and drone-trained GMM, ensuring all cross-view representations are embedded in a feature space induced solely by drone data.

Experimental Results

Quantitative Analysis

Extensive benchmarking is performed on two standard datasets: University-1652 and SUES-200, using Recall@1 (R@1) and Average Precision (AP) as primary metrics. The SFT framework substantially outperforms all protocol-matched satellite-free generalization baselines (e.g., AnyLoc, DINOv3, MambaVision). On University-1652, the framework achieves 62.05% R@1 (Drone$\rightarrow$Satellite) and 51.07% R@1 (Satellite$\rightarrow$Drone), representing relative gains of approximately 80% over the strongest generalization-only competitors.

On SUES-200, SFT outperforms all generalization baselines consistently across drone altitudes (150–300 m), with R@1 ranging from 78.75% to 93.50% for Drone$\rightarrow$Satellite retrieval. The absolute gap to supervised or satellite-trained methods remains, but is substantially narrowed compared to baseline models operating without satellite data.

Qualitative Insights

Qualitative retrieval analyses highlight that the 3DGS-based pseudo-orthophotos align well with satellite hypotheses, representing block shapes, road networks, and building outlines with high geometric fidelity. Failure cases predominantly occur in scenes with occlusions or repetitive layouts, suggesting directions for further improvement.

Ablation Studies

The results of a comprehensive ablation study demonstrate:

• 3DGS Reconstruction is essential; omitting it causes significant performance degeneration.
• Fisher Vector Aggregation with a drone-trained GMM outperforms VLAD and SoftVLAD, indicating the value of second-order statistics.
• Vocabulary Size Saturation is observed beyond 256 GMM components, and large descriptor sampling pools ensure stable estimation.
• LaMa Inpainting and geometry-guided completion critically contribute to downstream retrieval robustness.

Runtime

The majority of computation lies in offline geometry preprocessing and visual vocabulary learning; test-time feature extraction and retrieval are efficient and scalable.

Theoretical and Practical Implications

This work validates the feasibility of satellite-free training for DVGL, establishing a robust baseline that leverages explicit 3D geometric normalization to mitigate the cross-view domain gap. It demonstrates that physical scene representation (via neural, geometry-aligned approaches) and drone-only feature aggregation produce representations that successfully bridge to satellite imagery at test time, without satellite data or paired supervision during training.

The approach sets a re-usable pipeline for localization under satellite denial or restriction, with strong implications for autonomous navigation, surveillance, and security in adversarial or high-stakes operational scenarios.

Future Directions

Potential future developments include:

• Improving efficiency and scalability of 3D reconstruction for edge hardware
• Incorporating semantic and temporal priors to disambiguate repetitive or ambiguous scenes
• Leveraging multimodal context (e.g., language, maps) as structural side information
• Adapting SFT to broader cross-domain geo-localization, such as ground-to-aerial or urban-to-topography

Conclusion

The satellite-free DVGL pipeline presented in this paper, centered on 3DGS-based viewpoint normalization and drone-only Fisher vector aggregation, substantially outperforms existing generalization-only baselines and narrows the performance gap to satellite-trained methods. These findings establish SFT as a highly practical foundation for geo-localization in environments where satellite imagery is inaccessible during training, and provide a robust blueprint for further research in cross-view localization under real-world constraints (2604.01581).