Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

VFM-Loc: Zero-Shot Cross-View Geo-Localization via Aligning Discriminative Visual Hierarchies

Jun Lu, Zehao Sang, Haoqi Wei, Xiangyun Liu, Kun Zhu, Haitao Guo, Zhihui Gong, Lei Ding

Abstract

Cross-View Geo-Localization (CVGL) in remote sensing aims to locate a drone-view query by matching it to geo-tagged satellite images. Although supervised methods have achieved strong results on closeset benchmarks, they often fail to generalize to unconstrained, real-world scenarios due to severe viewpoint differences and dataset bias. To overcome these limitations, we present VFM-Loc, a training-free framework for zero-shot CVGL that leverages the generalizable visual representations from vision foundational models (VFMs). VFM-Loc identifies and matches discriminative visual clues across different viewpoints through a progressive alignment strategy. First, we design a hierarchical clue extraction mechanism using Generalized Mean pooling and Scale-Weighted RMAC to preserve distinctive visual clues across scales while maintaining hierarchical confidence. Second, we introduce a statistical manifold alignment pipeline based on domain-wise PCA and Orthogonal Procrustes analysis, linearly aligning heterogeneous feature distributions in a shared metric space. Experiments demonstrate that VFM-Loc exhibits strong zero-shot accuracy on standard benchmarks and surpasses supervised methods by over 20% in Recall@1 on the challenging LO-UCV dataset with large oblique angles. This work highlights that principled alignment of pre-trained features can effectively bridge the cross-view gap, establishing a robust and training-free paradigm for real-world CVGL. The relevant code is made available at: https://github.com/DingLei14/VFM-Loc.

VFM-Loc: Zero-Shot Cross-View Geo-Localization via Aligning Discriminative Visual Hierarchies

Abstract

Cross-View Geo-Localization (CVGL) in remote sensing aims to locate a drone-view query by matching it to geo-tagged satellite images. Although supervised methods have achieved strong results on closeset benchmarks, they often fail to generalize to unconstrained, real-world scenarios due to severe viewpoint differences and dataset bias. To overcome these limitations, we present VFM-Loc, a training-free framework for zero-shot CVGL that leverages the generalizable visual representations from vision foundational models (VFMs). VFM-Loc identifies and matches discriminative visual clues across different viewpoints through a progressive alignment strategy. First, we design a hierarchical clue extraction mechanism using Generalized Mean pooling and Scale-Weighted RMAC to preserve distinctive visual clues across scales while maintaining hierarchical confidence. Second, we introduce a statistical manifold alignment pipeline based on domain-wise PCA and Orthogonal Procrustes analysis, linearly aligning heterogeneous feature distributions in a shared metric space. Experiments demonstrate that VFM-Loc exhibits strong zero-shot accuracy on standard benchmarks and surpasses supervised methods by over 20% in Recall@1 on the challenging LO-UCV dataset with large oblique angles. This work highlights that principled alignment of pre-trained features can effectively bridge the cross-view gap, establishing a robust and training-free paradigm for real-world CVGL. The relevant code is made available at: https://github.com/DingLei14/VFM-Loc.
Paper Structure (21 sections, 10 equations, 6 figures, 8 tables)

This paper contains 21 sections, 10 equations, 6 figures, 8 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Cross-View Geo-Localization
  4. Unsupervised & Zero-Shot Learning
  5. Methodology
  6. Problem Formulation and System Overview
  7. Hierarchical Extraction of Discriminative Visual Clues
  8. Statistical Manifold Alignment
  9. Experiments
  10. CVGL Benchmarks and Evaluation Metrics
  11. Ablation Study
  12. Effectiveness of the proposed modules:
  13. Visualization on Manifold Alignment:
  14. Visualization of Similarity Responses
  15. Impact of $\alpha$ in R-MAC:
  16. ...and 6 more sections

Figures (6)

  • Figure 1: Key challenges of zero-shot CVGL: disparity in feature distribution. Drone and satellite visual manifolds suffer severe misalignment due to inherently different viewing geometry, thus direct zero-shot matching fails under viewpoint discrepancies.
  • Figure 2: The proposed VFM-Loc, a zero-shot CVGL framework, operates in three training-free stages: (1) Feature extraction using a VFM; (2) Hierarchical discriminative clue extraction via GeM pooling and scale-weighted R-MAC; (3) Statistical manifold alignment based on domain-wise PCA and orthogonal Procrustes rotation, projecting heterogeneous features into a unified embedding space for retrieval.
  • Figure 3: Sample image pairs from the constructed LO-UCV dataset. Top: satellite images; Bottom: high-resolution images from fixed-wing UAVs with large tilt (>45°).
  • Figure 4: t-SNE visualization of cross-view features before and after manifold alignment.
  • Figure 5: Similarity response visualization before and after manifold alignment.
  • ...and 1 more figures