Skip to content
Open access

SSAN: Structured Spatial Aggregation Network With Sparse Partition Routing for Cross-View Geo-Localization

2026 · IEEE Transactions on Geoscience and Remote Sensing · Vol 64, pp. 5646415-5646415 · 0 citations · 50 references

Abstract

Cross-view geo-localization (CVGL) aims to retrieve images depicting the same geographic target across heterogeneous viewpoints. Drone–satellite matching remains challenging because holistic descriptors may dilute localized geographic cues, predefined partitions are sensitive to viewpoint- and altitude-induced spatial variations, and exhaustive cross-view interaction may introduce unreliable associations between spatially misaligned regions. To address these issues, we propose a structured spatial aggregation network (SSAN), which couples response-driven regional aggregation with training-time sparse structural regularization. In particular, the soft partitioning module (SPM) learns continuous partition-response maps directly from retrieval supervision and aggregates reduced visual features into structured regional descriptors. Area-aware normalization (AAN) further compensates for differences in partition-response area, attenuating excessive dependence of regional descriptor magnitudes on response extent. Based on the learned response layouts, sparse partition routing (SPR) constructs compact regional tokens, estimates bidirectional cross-view affinities, and retains Top- $r$ high-affinity associations for structural refinement. SPR is applied only during training as a structural regularizer, preserving independent query–gallery encoding at retrieval. Extensive experiments on University-1652, SUES-200, and multiweather University-1652 demonstrate the effectiveness and robustness of SSAN under viewpoint, altitude, spatial-shift, and environmental variations.

Read PDF

We use cookies to run the site and, with your consent, for analytics and to show ads. See our Cookie Policy.