Abstract
Satellite-to-street-view image matching is essential for precise geo-localisation but remains challenging because of the substantial geometric and appearance differences between the two viewpoints. In addition, practical applications often involve registration errors, limited ground-view fields of view, and partial observation loss. To address these issues, this paper proposes SLCNet, a Scene-Robust Layout-Consistent Network for cross-view geo-localisation. The proposed method introduces a scene-robust training strategy with controlled satellite image misalignment and ground-view field-of-view occlusion to improve robustness under non-ideal input conditions. A Dual-Coordinate Gated Attention module enhances direction-sensitive structural features, while a layout consistency loss constrains intermediate layout representations across viewpoints to improve structural correspondence learning. Experimental results on the CVUSA, CVACT, and VIGOR datasets demonstrate that SLCNet achieves competitive matching performance and exhibits improved robustness under off-centre matching and various input perturbation settings.