S$^3$Geo: Structure-Semantic Synergistic Learning for Cross-View Geo-Localization

1 Introduction

• Existing visual approaches can overlook fine-grained structural differences and confuse visually similar regions with different semantic categories.Global features may miss discriminative local regions, while semantic priors are often absent.Source: §1 · pp. 1–2
• S3Geo achieves state-of-the-art performance on University-1652 and SUES-200 without extra inference overhead.It combines structural and semantic learning for cross-view matching.Source: §1 · p. 2
• S3Geo uses Decoupled Query Pooling with OT-based query-level contrastive learning to model local structure and soft cross-view matches.This is intended to address spatial misalignment between views.Source: §1 · p. 2
• Its CLIP-based Semantic Knowledge Distillation transfers language-aligned priors to filter deceptive hard negatives and support precise spatial correspondences.The framework unifies structural and semantic learning without additional annotations.Source: §1 · p. 2

2.1 Cross-view Geo-localization

• Global CNN and transformer representations capture visual context but can miss fine structural details and remain sensitive to cross-view misalignment.These limitations motivate region-level correspondence and semantic-consistency modeling.Source: §2.1 · p. 2

2.2 Structural Modeling for Cross-view Matching

• Local features, attention, multi-scale representations, and geometric aggregation have been explored, yet implicit alignment and strict layout assumptions impede reliable matching under drastic viewpoint changes.Naive one-to-one local matching can create incorrect associations when region ordering differs.Source: §2.2 · p. 2

2.3 Semantic Learning for Cross-view Matching

• Semantic approaches including SIGN and CLIP-UG add semantic cues, but feature-level fusion does not explicitly preserve relational structure for hard-negative discrimination.Figure 2 places feature- and relation-level CLIP distillation alongside global alignment and DQP structural matching.Source: §2.3 · p. 2, Figure 2 · p. 3

3.1 Backbone Representation and Global Alignment

• A shared DINO encoder and global pooling provide holistic drone and satellite features in a unified space, trained with bidirectional retrieval and location-ID supervision.The global features form the basis for instance-level matching.Source: §3.1 · p. 3

3.2 Decoupled Query Pooling Module

• DQP uses learnable queries and cross-attention to extract complementary region descriptors from dense patch tokens, addressing global pooling’s loss of local structure.Each query acts as a region probe that selectively aggregates informative tokens.Source: §3.2 · p. 4
• Because query order can differ across views, entropy-regularized OT supplies soft query correspondences, while symmetric contrastive supervision and attention decorrelation encourage complementary regions.The loss provides fine-grained structural supervision alongside global contrastive learning.Source: §3.2 · p. 4

3.3 Semantic Knowledge Distillation Strategy

• SKD transfers semantic priors from a frozen CLIP encoder by matching cross-view similarity distributions, preserving ranking structure and complementing structural modeling.The unified objective combines feature-level semantic alignment with relation-level ranking consistency.Source: §3.3 · pp. 4–5

3.4 Overall Training Objective

• Training combines global alignment, classification, DQP structural alignment, and CLIP distillation; inference uses only the global branch for similarity computation, without additional overhead.Source: §3.4 · p. 5

4.1 Datasets and Evaluation Metrics

• University-1652 covers drone, satellite, and ground views from 1,652 buildings; SUES-200 adds drone images at four altitudes, and both are evaluated with R@K and AP.The University-1652 evaluation uses Drone→Satellite and Satellite→Drone settings; SUES-200 spans 150m, 200m, 250m, and 300m flight heights.Source: §4.1 · p. 5

4.3 Comparison with State-of-the-Art Methods

• S3Geo outperforms all existing methods across University-1652 evaluation metrics and leads across SUES-200 altitude settings and retrieval directions.On University-1652, it reports 96.19% R@1 and 96.86% AP for Drone→Satellite, and 96.72% R@1 and 95.10% AP for Satellite→Drone; on SUES-200, Satellite→Drone R@1 is 100.00% at all altitude levels, while Drone→Satellite reaches 99.23% at 150m and 100.00% at 300m.Source: Table 1 · p. 5, §4.3 · pp. 5–6, Table 2 · p. 6

4.2 Implementation Details

• Experiments use a DINOv3 backbone with a frozen CLIP visual teacher, 512 × 512 inputs, and four learnable DQP queries.All but the last two DINOv3 transformer blocks are frozen during fine-tuning; the stated training setup uses AdamW and batch size 32.Source: §4.2 · p. 5

4.4 Cross-Region Transferability

• In zero-shot transfer from University-1652 to SUES-200 without fine-tuning, S3Geo achieves the best R@1 and AP across altitudes in both retrieval directions.The evaluation tests generalization to an unseen environment without target-domain adaptation.Source: §4.4 · pp. 6–7, Table 3 · p. 7

4.5 Ablation Study

• Ablations find that DQP and SKD each improve the global baseline, their combination performs best, and OT alignment improves on naive matching.Feature-level and relation-level distillation each improve performance, and combining them further enhances it; the authors describe structural modeling and semantic enhancement as complementary.Source: §4.5 · p. 7
• With DINOv2, S3Geo improves R@1 by 2.09 and 1.11 percentage points for Drone→Satellite and Satellite→Drone, respectively; tested semantic teachers differ by less than 0.2 R@1 points.The reported teacher comparisons replace CLIP with SigLIP or RemoteCLIP.Source: §4.5 · p. 7, Table 5 · p. 8
• Performance peaks at M=4 queries; increasing the count further brings marginal gains or slight degradation, while reasonable loss-weight changes cause only minor fluctuations.The authors attribute possible degradation at higher query counts to overlapping attention regions and redundancy, and report the best loss-weight setting as λd=0.5 and λs=0.5.Source: §4.5 · pp. 7–8, Table 6 · p. 8, Table 7 · p. 8

4.6 Visualization Analysis

• S3Geo produces more concentrated and consistent attention patterns than the baseline, whose heatmaps already capture similar salient regions.The authors suggest these refinements suppress noise and enhance meaningful structural information; they describe the visual differences as relatively subtle.Source: §4.6 · p. 8, Figure 3 · p. 8

• S3Geo combines Decoupled Query Pooling with OT-based soft alignment and CLIP-based semantic distillation to capture region-level correspondences and improve discrimination of hard negatives.Source: §5 · p. 8
• Experiments on University-1652 and SUES-200 demonstrate consistent performance gains and strong generalization ability.Source: §5 · p. 8
• The results highlight the importance of jointly modeling structural and semantic information for robust cross-view representations.Source: §5 · p. 8