Xu Liu

Remote-sensing visual localization aims to accurately identify target regions from natural-language descriptions under complex imaging conditions. Although recent multimodal approaches have achieved notable progress via text-guided attention mechanisms, their decision processes often treat text-induced correlations as the primary source of evidence, resulting in limited robustness under distribution shifts. From a causal perspective, visual grounding can be viewed as a perception–decision process influenced by latent causal factors. Specifically, image observations are jointly affected by stable structural causal factors and external confounders, among which the former serve as the primary source of reliable and transferable evidence for prediction. Motivated by this insight, we propose a causal prior–driven visual grounding (CPVG) framework that explicitly models and exploits structural causal priors in visual data. CPVG instantiates causal prior–conditioned query modeling (CPCQM) to build a dual-query mechanism where structural causal priors and textual information cooperate, enabling priors to directly participate in query generation and visual evidence retrieval. Furthermore, to assess whether the performance gains brought by prior-guided queries are associated with more causally relevant information, we propose causal evidence attribution learning (CEAL), which introduces controlled counterfactual comparisons to encourage the model to rely more on stable structural evidence. Experiments conducted on the DIOR-RSVG and OPT-RSVG benchmarks demonstrate that CPVG yields consistent improvements in both localization accuracy and robustness, validating the effectiveness of combining distributed causal-prior modeling with causal evidence attribution learning.

Recent advancements in natural image visual grounding have shown promising results. However, existing approaches struggle with precise grounding in remote sensing (RS) images. This challenge arises from two key issues. First, the prevalence of small objects and dense spatial distributions in RS images makes accurate localization difficult. Second, the large domain gap and lack of contextual priors in RS data hinder language-vision alignment. In this paper, we propose RefineVG, a novel framework for fine-grained visual grounding in high-resolution RS images. Unlike methods that require training separate models for coarse localization and fine-grained refinement, RefineVG employs a unified model that iteratively refines predictions. Specifically, RefineVG first predicts a coarse bounding box using a multi-scale cross-modality encoder and a Transformer-based decoder, which progressively align and fuse visual and textual feature. Then, the predicted region is enlarged to capture additional contextual feature. Combined with only category information as text, the cropped region is then re-encoded by the same model to obtain a refined grounding result. Finally, a fusion block combines the coarse and refined predictions. To support this process, we propose a data augmentation strategy by randomly expanding ground truth bounding boxes in training dataset, simulating spatial uncertainty.Extensive experiment results show that RefineVG achieves outstanding performance on three benchmarks, outperforming the second-best method by 5.11% on the DIOR-RSVG dataset, 8.61% on the OPT-RSVG dataset, and 3.61% on the RSVG-HR dataset in P@0.9. Our code are available at https://github.com/XiaowZhang00/RefineVG.

Remote sensing visual grounding (RSVG) models commonly rely on backbones pretrained on natural image datasets. However, the inherent domain gap between natural images and remote sensing imagery significantly degrades model transferability, particularly when RSVG training samples are limited. Existing approaches primarily focus on enhancing localization accuracy through multi-scale visual–textual alignment and process each image–text pair independently, which limits their generalization capability under constrained supervision. In this work, we propose VisproVG, a novel framework that explicitly incorporates category-aware visual priors into the linguistic stream to mitigate domain shift and data scarcity. Specifically, VisproVG introduces a sample-level visual prompting mechanism that retrieves instance-level visual cues from a dynamically maintained, category-specific memory bank, which serve as stable and interpretable visual priors. To further improve semantic consistency across feature scales, we incorporate a lightweight cross-scale semantic refinement module that performs hierarchical top–down refinement. Beyond architectural design, we conduct a systematic analysis of the relationship between decoder confidence and localization accuracy. The analysis indicates that low-confidence predictions, while accounting for only a small fraction of test samples, exhibit substantially greater potential for improvement. Motivated by this observation, we introduce a confidence-guided adaptive test-time augmentation (TTA) strategy that selectively applies augmentation to uncertain samples, achieving a favorable balance between accuracy gains and computational efficiency. Extensive experiments on three RSVG benchmarks demonstrate the effectiveness of the proposed VisproVG. Notably, on the small-scale RSVG-HR dataset, VisproVG surpasses the second-best method by 5.05% meanIoU, highlighting its robustness under limited training data. Our code are available at https://github.com/XiaowZhang00/VisproVG.