Open Access

ARTICLE

Cross-View Geo-Localization via Dynamic Multi-Positive Mining from Unlabeled Data

Long Yu^1,2,3, Ma Zhu^1,2,3,*, Xu Wang^1,2,3, Yang Pei^1,2,3, Chunfang Yang^1,2,3

1 Information Engineering University, Zhengzhou, China
2 Key Laboratory of Cyberspace Security, Ministry of Education of China, Zhengzhou, China
3 Henan Key Laboratory of Cyberspace Situation Awareness, Zhengzhou, China

* Corresponding Author: Ma Zhu. Email:

Computers, Materials & Continua 2026, 89(2), 94 https://doi.org/10.32604/cmc.2026.087636

Received 20 June 2026; Accepted 19 August 2026; Issue published 15 September 2026

Abstract

Cross-view geo-localization (CVGL) estimates the location of a street-level image by retrieving its matching GPS-tagged satellite tile. Semi-supervised methods reduce the need for dense annotations by mining pseudo labels, but most of them keep only one positive reference for each query. In real-world galleries, several overlapping satellite tiles may cover the same ground location. As a result, valid matches can be discarded as negatives, which gives the model conflicting supervision. To address this problem, we propose DMP-Geo, a semi-supervised cross-view geo-localization method that mines multiple positives for each query from unlabeled data. A bird’s-eye fusion encoder is designed to combine each panorama with its bird’s-eye-view projection, thereby reducing the viewpoint gap between ground and satellite images. A breakpoint-based mining strategy is then proposed to retain multiple top-ranked candidates before the first clear similarity drop and verify them through mutual consistency. Finally, a confidence-weighted multi-positive contrastive loss is introduced to assign different weights to the mined positives according to their confidence. With only 10%–30% of the annotations, DMP-Geo consistently outperforms the existing semi-supervised baseline on CVUSA, CVACT, and VIGOR, and demonstrates strong generalization to unseen cities.

---

Keywords

---