acceptodds
Under review as a conference paper at ICLR 2027

GrabVG: Graph-Attentive Binding for Visual Grounding in UAV Imagery

Abstract

Visual grounding in Unmanned Aerial Vehicle (UAV) imagery aims to localize a target object in complex bird's-eye-view scenes according to a natural language description. However, the abundance of small, densely distributed, and visually similar objects creates high visual redundancy, while repetitive local configurations give rise to strong topological ambiguity. Existing approaches mainly focus on visual–language feature alignment or dense contextual interaction, yet they struggle to distinguish subtle inter-instance differences and effectively exploit spatial topological structures, leading to inaccurate grounding in highly crowded scenarios. To address these challenges, we propose GrabVG, a novel visual grounding framework inspired by human visual search. GrabVG explicitly decomposes grounding into two sequential stages: preattentive hypothesis search and graph-attentive feature binding. Specifically, we first generate a compact set of reliable object hypotheses through distillation-guided proposal induction and text-aware hypothesis filtering, substantially reducing background distractions and semantic mismatches. These hypotheses are then organized into a sparse graph, where language-guided intra-instance visual cues and inter-instance topological relationships are jointly bound and propagated via graph attention, enabling efficient spatial reasoning and accurate target localization. Extensive experiments on AerialVG and AerialSense show that GrabVG achieves a favorable accuracy–speed trade-off, reaching 67.31% and 80.34% [email protected] and outperforming the corresponding baselines by 10.55 and 8.76 percentage points, respectively. The code will be made publicly available.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
UAV Visual Grounding

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_rt6SOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes the GrabVG, a visual grounding framework inspired by human visual search. In its scheme, it explicitly decomposes a grounding task into two sequential stages: preattentive hypothesis search and graph-attentive feature binding. Moreover, it generates a compact set of reliable object hypotheses through distillation-guided proposal induction and text-aware hypothesis filtering to reduce background distractions and semantic mismatches. Extensive experiments on AerialVG and AerialSense are reported to demonstrate the effectiveness and efficiency of proposed GrabVG framework.

Critical strengths

There are some Critical Strengths in this paper.

  1. It reformulates a crowded UAV visual grounding task as a search-before-binding problem, proposing a visual grounding framework GrabVG.
  2. Two strategies: preattentive hypothesis search and graph-attentive feature binding, are designed to enhance the visual grounding performance of GrabVG framework.
  3. Extensive experiments on two public datasets: AerialVG Liu et al. (2025) and AerialSense Guo et al. (2025) , are performed to demonstrate the effectivenss and superiority of proposed GrabVG framework.
Critical weaknesses

Although this paper has some highlights, it has some Critical Weaknesses.

  1. This paper is with limited novelty. The pimary ideas, including preattentive hypothesis search and graph-attentive feature binding, have been similarily explored in existing methods, e.g., [1] RCCF: Recursive Content-Candidate Filtering for Robust Visual Grounding, [2] QueryRec: Two-Stage Query refinement for Visual Grounding, [3] GraphDET: Graph-Attentive Topology Reasoning for UAV Small Object Detection, and [4] VR-GAT: Visual Relationship Graph Attention Transformer. Inspired by these methods, this paper uses the similar ideas to specially constuct a new framework for UAV Imagery scenes, limiting its orignal novelty and contributions.
  2. In the explanation to GrabVG framework, the definitions of some loss terms are missed, e.g., the L_filt and L. Moreover, for easily understanding, in Fig 2(a), it is better to briefly show how to compute its training loss.
  3. In experiments, e.g., Table 1 and Table 2, too few methods (especially 2026-published methods) are used for performance comparisons. In table 2, even no 2026-published method is compared. Moreover, only IoU metric is used for performance evaluation. Other metrics, e.g., mAP could be adopted? In Table 4, only Inference-speed is presented. For deep learning models, parameter scale and FLOPs are two usual metrics.
  4. The viualization evidences to demonstrate the effectivenss and superority of GrabVG are unsufficient. Only resutant evidences are demonstrated in experiments (including in Supplementary Material).
Questions for authors

I have some concerns for this paper. They are listed as follows. Please further clarify or explain them.

1.This paper is with limited novelty.

The pimary ideas, including preattentive hypothesis search and graph-attentive feature binding, have been similarily explored in existing methods, e.g., [1] RCCF: Recursive Content-Candidate Filtering for Robust Visual Grounding, [2] QueryRec: Two-Stage Query refinement for Visual Grounding, [3] GraphDET: Graph-Attentive Topology Reasoning for UAV Small Object Detection, and [4] VR-GAT: Visual Relationship Graph Attention Transformer. Inspired by these methods, this paper uses the similar ideas to specially constuct a new framework for UAV Imagery scenes, limiting its novelty and contributions.

2.Loss definition and presentation

In the explanation to GrabVG framework, the definitions of some loss terms are missed, e.g., the L_filt and L. Moreover, for easily understanding, in Fig 2(a), it is better to briefly show how to compute its training loss.

3.Compared methods and metrics

In experiments, e.g., Table 1 and Table 2, too few methods (especially 2026-published methods) are used for performance comparisons. In table 2, even no 2026-published method is compared. I wonder if these methods have truly contained the current SOTA ones in visual gounding field of UAV Imagery. Moreover, only IoU metric is used for performance evaluation. Other metrics, e.g., mAP could be adopted? In Table 4, only Inference-speed is presented. For deep learning models, parameter scale and FLOPs are two usual metrics.

4.Viualization evidences

The viualization evidences to demonstrate the effectivenss and superority of GrabVG are unsufficient. Only resutant evidences are demonstrated in experiments (including in Supplementary Material).

Reviewer_LGtpOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies visual grounding in UAV imagery, where many small, similar objects create visual redundancy and topological ambiguity. Existing methods either fail to capture fine-grained differences or introduce noise through dense relation modeling. Inspired by human visual search, the authors propose GrabVG, a two-stage framework. The first stage named Preattentive Hypothesis Search uses expression-aligned pseudo-labels from teacher models to train a proposal generator, followed by lightweight text-aware filtering to keep the top-K candidates. The second stage is Graph-Attentive Feature Binding, which combines two parts: intra-instance appearance binding uses learnable text-conditioned anchors to sample multi-scale local features, while inter-instance topology binding applies language-guided graph attention over a sparse k-NN graph to model spatial relationships. Finally, a contrastive scoring module selects the target box. On AerialVG and AerialSense, GrabVG reaches 67.31% and 80.34% [email protected], improving over the strongest baselines by 10.55 and 8.76 points while running at 28.51 FPS.

Critical strengths
  1. The paper clearly identifies the key challenges of UAV visual grounding and proposes a coherent two-stage framework inspired by human visual search, where each component is specifically designed to address the corresponding challenges.
  2. The proposed GrabVG achieves substantial improvements over existing methods on both AerialVG and AerialSense, outperforming the strongest baselines by 10.55 and 8.76 percentage points in [email protected], respectively. Moreover, the method maintains a competitive inference speed of 28.51 FPS, demonstrating a favorable trade-off between accuracy and efficiency.
Critical weaknesses

The source of the improvement is not yet clearly attributed to the proposed binding design. On AerialVG, adding the teacher pseudo-annotations alone raises Top-1 from 56.76% to 62.84%. Filtering and both binding modules together add a further 4.47 points, and the filter contributes only 0.30. The baselines in Tables 1 and 2 do not receive this extra supervision. On AerialSense, the teacher choice changes [email protected] by 16 points (64.04 vs 80.34), more than any architectural component. The current experiments therefore do not establish that graph-attentive feature binding, rather than improved proposal supervision, is the main source of the claimed gain.

Questions for authors
  1. The GT referent is appended when no teacher prediction matches it. How often does this happen for each teacher? What fraction of the annotated boxes in AerialVG do the teacher pseudo-boxes recover, and how does proposal induction perform without the fallback?
  2. The kk-NN graph is built from box centers only, does not depend on the expression, and stays fixed because boxes are never updated. When the reference object is not among the 6 nearest candidates, how is the relation resolved? Can you report results for several values of kk, and accuracy binned by referent–reference distance?
  3. Can you report accuracy by object density, target size and relation type, compared with PropVG*, to show that the gains concentrate on the cases GrabVG is designed for?
  4. AerialSense has no validation split. Were the teacher source and other design choices selected on the test split? Were the baselines in Table 2 retrained on your 75%/25% split with the same backbone and resolution?
Reviewer_unzoOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes GrabVG to address the visual grounding problem in crowded drone imagery. The method combines pseudo-label supervision, text-aware candidate filtering, language-guided local sampling, and sparse geometric graph attention. Finally, the CRS module from PropVG is employed for final selection. The paper reports 67.31% and 80.34% [email protected] on AerialVG and AerialSense, respectively. The authors also conducted ablation experiments on individual components, graph topologies and pseudo-annotation sources.

Critical strengths
  1. Clear motivation. The separation of candidate screening and appearance–topology reasoning is well motivated for crowded UAV scenes.

  2. Strong reported results. GrabVG achieves strong grounding performance on AerialVG and AerialSense with competitive inference speed.

  3. Comprehensive ablations. The experiments evaluate individual components, graph topologies and pseudo-annotation sources.

  4. Useful qualitative analysis. Visualizations illustrate how appearance and spatial relations help resolve ambiguities, while supplementary details support reproducibility.

Critical weaknesses
  1. Missing Top-5 results on AerialSense. Section 4.1 defines Top-5, but Tables 2 and 10 omit it without explanation. Please report this metric or justify its omission.

  2. Incomplete hyperparameter analysis. The paper do not study the retention rate, neighbor count kk, anchor count MM, or loss weights. Sensitivity analyses of these hyperparameter are necessary.

  3. Incomplete comparison with recent methods. The related-work section discusses recent approaches such as ProVG and GeoVis, but neither appears in the main quantitative comparison. If these methods are applicable to the same task and evaluation protocol, direct comparisons would strengthen the state-of-the-art claim. Otherwise, the authors should explain the differences in task or evaluation settings that prevent a fair comparison.

  4. Unclear attribution of performance gains. Tables 3 and 9 provide matched-supervision evidence on AerialVG. However, Table 10 shows that changing the pseudo-annotation source shifts AerialSense [email protected] from 64.04% to 80.34%, while an equivalent component breakdown is not reported for this dataset. Clarifying the architectural gain under a fixed teacher, together with the teacher configurations, would strengthen the attribution.

  5. Limited evidence for learned filtering. In Table 5, the NMS variant achieves 67.25% [email protected] and 90.03% Top-5 accuracy, compared with 67.31% and 89.43% for the learned filtering variant. Thus, the learned filter offers only a marginal Top-1 improvement while producing lower Top-5 accuracy in this comparison. A clearer discussion of its practical advantage, including its contribution to inference efficiency, would strengthen the design rationale.

Questions for authors

Please refer to weakness

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.