Cross-view object geo-localization (CVOGL) aims to localize a target in a satellite image from an observation captured under another viewpoint. Existing methods typically specify the query with a click point or bounding box, marking the target region but underusing the semantic attributes and relations conveyed by natural language. Moreover, referring expressions are often viewpoint-dependent, making query-view descriptions difficult to ground directly in the satellite image. In this paper, we formulate cross-view visual grounding (CVVG), extending conventional point- or box-prompted CVOGL to instance grounding jointly guided by a referring expression and a visual exemplar. We present GAV (Grounding Across Views), a view-aware framework for selectively transferring visual and linguistic evidence across views. Its core is view-aware referring reasoning: target-centered descriptions are derived from the marked query image, checked for target-category consistency, and augmented with a direction-free role cue. Exemplar prompting, cross-view alignment, and scene-level discrimination ground this semantic evidence in a satellite instance. GAV achieves a new state of the art on MoP-UAV with 62.55% [email protected]. We also introduce CVOGL-Text, a benchmark for text-guided grounding from street- and drone-view queries to satellite images. GAV outperforms the click-point baseline by 7.09 and 6.79 [email protected] points on the street- and drone-view splits, respectively, demonstrating robustness to larger viewpoint gaps. The code and annotations will be publicly released.
Figures & tables
Figure 2: Architecture of GAV. Lower panels, left to right: (A) View-Aware Referring Reasoning, (B) Multi-Token Cross-View Exemplar Prompting, (C) Cross-View Contrastive Alignment, and (D) Scene-Level Discriminative Grounding. B includes exemplar token encoding and gated residual cross-attention. C learns projection heads; D adds scene-ranking supervision and reuses these heads for verification without new parameters. A runs offline; its speech-bubble input denotes the factual instruction, while source T is reserved for checking/fallback. Satellite ground-truth boxes are never inference inputs.
Figure 3: View-Aware Referring Reasoning (A) and its effect on grounding. Top: a frozen VLM samples descriptions from the marked query image and a factual instruction; content-word medoid selection, rule-based sanitization, target-category checking, and a direction-free query-frame role produce Tsat . Source T supplies the category check and conditional fallback, not the generative prompt. Bottom: the same GAV weights, satellite image, and exemplar yield IoU 0.512 with T and 0.900 with Tsat on this sample ( λ=0 ). Satellite annotations and target crops illustrate correspondence only. This paired example does not measure an aggregate consistency score.
GeoText -1652
CVOGL
MoP -UAV
MoP -UAV †
CVOGL -Text †
Task
IGL
OGL
OGL
OGL
OGL
Query view
UAV
drone, street
UAV
UAV
drone, street
Point
–
✓
✓
✓
✓
Box
✓
–
✓
✓
–
Language
template
–
template
derived
derived
Categories
–
19
–
–
19
Table 1: Comparison with cross-view geo-localization benchmarks. IGL and OGL denote image- and object-level geo-localization, respectively. Columns marked † are contributed by this work, and GeoText-1652 [ 4 ] counts follow Zhang et al. [36] .
Table 2: Comparison with state-of-the-art methods on the MoP-UAV test set. Baselines are reported by Zhang et al. [36] on the same split. GAV uses text and a query-view exemplar without a point prompt; its re-ranking weight is selected on held-out training data.
Table 3: Comparison on the CVOGL-Text test sets, with 973 samples each for street-view and drone-view queries. GAV uses derived expressions and Blocks A–D; λ=0.7 is selected on the validation split. Text-only and partial configurations provide input controls; dashes denote unreported mIoU.
Figure 4: Stage-wise qualitative visualization on MoP-UAV. Per-sample IoU curves are paired with the expression, exemplar-marked query image, ground truth, and cumulative predictions. The tinted column identifies the stage that corrects each case. The unchanged plot retains its original lettering: exemplar A, alignment B, and reasoning C correspond to Blocks B, C, and A in the present organization.
Figure 5: Joint t-SNE visualization of frozen BERT embeddings for 5,000 paired original and derived expressions from the MoP-UAV test set, before role-cue concatenation. Both panels share one embedding space; gray points show the other corpus. Colors distinguish direction-anchored openers, center-anchored openers, and the remaining target-description group. Arrows connect six selected before/after pairs.
Table 4: Component ablation on the MoP-UAV test set. Rows add components to the fine-tuned text-only base; check marks indicate active components. Exem./Multi are the exemplar prompt and multi-token variant in B; Align is C; Reas./Pos. are expression derivation and the role cue in A; Scene/ReR are D1/D2. Each configuration uses one fixed-seed run.
Exemplar token count, with alignment (Blocks B + C)
Table 5: Exemplar-prompt design ablations on the MoP-UAV test set. Fusion variants use the text-only base; token-count variants keep cross-view alignment fixed.
Base configuration
λ∗
Standard
Re-ranked
Δ
Visual prompts (B + C)
0.7
0.5159
0.5364
+2.05
+ reasoning (A)
0.7
0.5796
0.5888
+0.92
+ scene supervision (D)
0.5
0.6209
0.6255
+0.46
Table 6: Alignment re-ranking ablations on the MoP-UAV test set. Standard and re-ranked [email protected] use identical weights; Δ is the gain in percentage points. Each λ∗ is selected on held-out training data.
State Key Laboratory of Multimodal Artificial Intelligence Systems, Institute of Automation, Chinese Academy of Sciences, and also with the School of Artificial Intelligence, University of Chinese Academy of Sciences · Beijing Aerospace Control Center