From Coarse to Fine: A Matching and Alignment Framework for Unsupervised Cross-View Geo-Localization

Xueyi Wang, Lele Zhang, Zheng Fan, Yang Liu, Chen Chen, Fang DengAAAI 2025 · DOI

TL;DR

This paper introduces a novel unsupervised framework for cross-view geo-localization, enabling accurate matching of satellite and drone imagery without requiring labeled data. By employing a coarse-to-fine approach that leverages cluster-based contrastive learning and an intermediate state transition module, the method achieves performance comparable to supervised techniques.

What Problem Does This Paper Address?

Cross-view geo-localization aims to determine the geographic location of an image by matching it with reference images from a different perspective, such as matching a drone image with a satellite image. Traditional methods often rely on supervised learning, requiring extensive labeled data or matched image pairs, which is costly and not always available. Existing unsupervised methods struggle with the significant feature discrepancies between different views, leading to suboptimal performance. This paper tackles the challenge of achieving robust and accurate cross-view geo-localization in an unsupervised manner, bridging the gap between diverse image modalities without relying on explicit annotations.

Methods

The proposed framework operates in a coarse-to-fine manner, integrating three key components:

  • Cluster-Based Contrastive Learning Baseline: This module initializes the process by applying DBSCAN clustering to features extracted from both satellite and drone views independently. Pseudo-labels are assigned based on these clusters, and cluster centroids are computed to form memory banks. This establishes an initial understanding of intra-view relationships.
  • Cross-View Location Alignment Module: This is the core of the method, designed to bridge the gap between the two views. It employs a novel "local-to-global" matching strategy that interactively matches individual instances to global cluster centroids, rather than relying solely on instance-to-instance or cluster-to-cluster matching. This approach is more robust to noisy samples. It then uses cross-view cluster contrastive learning to learn view-invariant, location-specific features by pulling query instances closer to their corresponding cross-view cluster centroids.
  • Intermediate State Transition Module: Recognizing the significant feature discrepancies between drone and satellite views, this module introduces a non-learning approach to create an "intermediate domain." It performs color transfer from satellite images to drone images in the Lab color space, aiming to reduce visual differences. This intermediate state then facilitates more effective cross-view contrastive learning by allowing both views to align within a common, less disparate space.

The overall optimization objective integrates losses from these three components.

Datasets

The proposed method is evaluated on two widely-used cross-view geo-localization datasets:

  • University-1652: A large-scale benchmark with 1,652 buildings across 72 universities, featuring drone and satellite images.
  • SUES-200: A multi-scene, multi-height dataset with drone and satellite images captured from various altitudes.

Key Insights and Ideas

  • Unsupervised is Achievable: The paper demonstrates that high-performance cross-view geo-localization is achievable without any labeled data, significantly reducing the reliance on costly annotation efforts.
  • Coarse-to-Fine Strategy: Moving from coarse cluster-level matching to fine instance-level alignment, and incorporating an intermediate state, proves crucial for handling the inherent discrepancies between views.
  • Local-to-Global Matching: The novel matching strategy that combines instance and cluster information is more robust than traditional instance-only or cluster-only approaches, effectively mitigating the impact of noisy samples.
  • Intermediate Domain for Alignment: The color transfer-based intermediate state transition module is a clever way to create a more unified feature space, making cross-view alignment more effective.
  • Bridging the Gap: The framework successfully bridges the significant visual and feature gaps between drone and satellite imagery, a long-standing challenge in the field.

Key Results

The proposed unsupervised method achieves state-of-the-art performance among unsupervised approaches and demonstrates comparable results to supervised methods on both University-1652 and SUES-200 datasets. For instance, on University-1652, the method achieves an R@1 score of 85.95 for drone-to-satellite and 90.33 for satellite-to-drone, outperforming the previous unsupervised SOTA by approximately 15-20%. Ablation studies confirm the effectiveness of each proposed module (Cross-View Location Alignment, Intermediate State Transition, and Local-to-Global Matching).

Limitations and Future Work

  • Clustering and Matching Errors: The authors acknowledge that errors in clustering and matching can still occur, potentially impacting performance, especially when data discrepancies are severe.
  • Data Discrepancy: While the intermediate state transition module helps, significant data discrepancies can still pose a challenge.
  • Future Work: The paper suggests that further exploration into more sophisticated intermediate state generation techniques and advanced clustering algorithms could lead to even better performance. Investigating the application of this framework to other cross-view matching scenarios beyond drone-satellite pairs is also a promising direction.

Summary updated

BibTeX
@inproceedings{wang2025from,
  author = {Xueyi Wang and Lele Zhang and Zheng Fan and Yang Liu and Chen Chen and Fang Deng},
  title = {{From Coarse to Fine: A Matching and Alignment Framework for Unsupervised Cross-View Geo-Localization}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2025},
  doi = {10.1609/aaai.v39i8.32865},
}
Figure