M 3 E: An observation-support-aware enhancement for BEV matching in fine-grained cross-view localization
Cross-view localization estimates the Three Degrees of Freedom (3-DoF) pose of a ground-view image with respect to a satellite image and is relevant when GNSS signals may be limited or unavailable. Recent Bird’s-Eye-View (BEV) matching-based methods have improved localization accuracy by transforming cross-view images into a shared top-down representation and estimating pose through feature matching. However, ground-view BEV features are constructed through view transformation and spatial aggregation rather than directly from observations, which introduces non-uniform observation support and may produce unreliable correspondences in weakly supported regions. We propose M 3E, an observation-support-aware enhancement for BEV matching in cross-view localization. M 3E improves the matching process from both global and local perspectives. The Geometry Consistency-Guided Optimization (GCO) scheme adaptively regulates feature-level supervision using Virtual Correspondence Error (VCE) as a geometry-consistency proxy for the current cross-view pose-estimation state. The Visibility-Aware Correspondence Selection (VCS) strategy uses a camera-distance observation-support prior to reweight matching candidates, encouraging pose estimation to rely more on regions better supported by ground-view observations. Rather than replacing existing pipelines, M 3E complements BEV matching-based methods and can be integrated into the evaluated learnable-BEV matching pipelines to improve correspondence reliability and pose estimation. Experiments on VIGOR, DReSS-D, and KITTI-CVL show that the most consistent benefit of M 3E lies in translation accuracy and robustness across the evaluated FG 2 and ViewBridge pipelines, while orientation gains are setting-dependent and less uniform, particularly in limited-FoV settings. The translation gains are especially evident in challenging unknown-orientation scenarios. The experimental results support the importance of spatially non-uniform observation reliability in BEV matching-based cross-view localization. M 3E introduces no additional learnable parameters and incurs negligible measured inference-time latency overhead under the tested settings, providing a practical enhancement for the evaluated learnable-BEV matching pipelines and offering a useful direction for more reliable correspondence reasoning.
Authors
- Xudong Lu (ORCID: https://orcid.org/0009-0008-9082-4084)
- Yi Wan (ORCID: https://orcid.org/0000-0001-6777-6047)
- Mingtao Xiong
- Panwang Xia (ORCID: https://orcid.org/0009-0005-1857-4437)
- Dong Wei (ORCID: https://orcid.org/0000-0001-8199-4243)
- Yifei Wang
- Junjian Zhang
- Hongwei Hu
- Zhi Zheng
- Xinyi Liu
- Yongjun Zhang
- Yi Liu
- Lei Yu
- Qiong Wu (ORCID: https://orcid.org/0009-0008-6213-9362)
Institutions
- Hong Kong Polytechnic University (HK)
- Chinese University of Hong Kong (HK)
- Ministry of Natural Resources (CN)
- Wuhan University (CN)
- Ant Group (China) (CN)
Publication Details
- Journal
- Geo-spatial Information Science
- Published
- 2026-11-17
- DOI
- https://doi.org/10.1080/10095020.2026.2757300
- Primary Topic
- Advanced Image and Video Retrieval Techniques
- Type
- article
- Field-Weighted Citation Impact
- 0.00
Funders
- Ant Group
- National Natural Science Foundation of China