Pith. sign in

REVIEW 5 major objections 5 minor 44 references

Transformer-based Spatial Grounding: A Comprehensive Survey

T0 review · 5 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read The paper claims to be the first comprehensive systematic review of transformer-based spatial grounding in images, and reports that TransVG, the RefCOCO dataset family, and IoU are the field's most-used model, datasets, and evaluation…

desk verdict A useful qualitative map of transformer-based spatial grounding, but the quantitative synthesis is unreproducible due to inconsistent corpus counts and a missing study list. read the letter →

arxiv 2507.12739 v1 pith:G4D34MAT submitted 2025-07-17 cs.CV cs.AI

Ijazul Haq , Muhammad Saqib , Yingjie Zhang This is my paper
classification cs.CVcs.AI
keywords spatialgroundingvisualreferringexpressioncomprehensiontransformerarchitecturessystematicliteraturereviewevaluationmetricsvision-languagemodelsdatasets
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

Spatial grounding links natural-language expressions to image regions, and the paper's central claim is that transformer-based work in this area now forms a measurable field with clear conventions. The authors present a systematic literature review covering 2018 to 2025, screening 523 retrieved articles down to 45 primary studies through a staged process, and then counting what those studies actually use. On their counts, TransVG is the most frequently adopted architecture (about 13 percent of papers), the RefCOCO series dominates dataset usage, and IoU is the standard spatial metric (around 65 percent). The survey also maps industrial application domains, finding heavy concentration in general grounding and emerging but underrepresented work in remote sensing, dialogue, and construction safety. A sympathetic reader would take the contribution as a structured, data-driven map of a young field, meant to guide dataset, architecture, and metric choices.

What carries the argument

The load-bearing mechanism is the systematic literature review pipeline itself. The authors build a Boolean query from a PICO (Population, Intervention, Comparison, Outcome) term grid, run it against five digital libraries, then reduce the retrieved articles through a Tollgate-style multi-stage screen (title/abstract, introduction/conclusion, full text), score survivors on a quality checklist with A/B/C tiers, and exclude studies scoring below 50 percent. From the remaining 45 primary studies they extract six dimensions per paper, namely model architecture, grounding task, datasets, metrics, techniques, and outcomes, and turn those extractions into frequency distributions and overlap analyses. The review's conclusions are carried entirely by this pipeline and the counts it produces.

What would settle it

Run the same Boolean query across the five databases with the stated inclusion and exclusion criteria and compare the resulting corpus: if the set of primary studies cannot be reconstructed without ambiguity, or if an independent screening yields materially different usage percentages for IoU, RefCOCO, or TransVG, then the reported frequencies are artifacts of corpus construction rather than stable facts about the field.

Watch

Extended reading notes

Core claim

On the paper's own terms, the finding is that transformer-based spatial grounding from 2018 to 2025 is not an unstructured collection of isolated models but a field with stable conventions. The authors find that most studies pair a CNN or vision-transformer visual backbone with a transformer language encoder, that TransVG serves as the most common reference architecture and is followed by MDETR, that RefCOCO, RefCOCO+, and RefCOCOg anchor evaluation while Flickr30K and ReferItGame appear frequently, and that IoU and accuracy dominate performance reporting while language metrics such as BLEU and METEOR appear mainly in generation and captioning tasks. They further find that bounding-box annotations dominate the benchmarks, that publication growth steepens sharply after 2021, and that industrial evaluation is uneven: general-purpose grounding is well covered, while safety-critical and geospatial domains use specialized datasets rarely and report inconsistent metrics. The paper presents these as evidence-based trends from a systematic corpus, not as theoretical results.

Load-bearing premise

The review's trend claims rest on the assumption that the 45 selected studies faithfully represent all transformer-based spatial grounding research from 2018 to 2025, yet the paper gives no included-study list and reports conflicting corpus counts (60, 45, and 46), so that representativeness is not yet demonstrated.

Editorial extensions

If this is right

  • A newcomer can benchmark against the field's de facto standard: a TransVG-style ViT-plus-BERT model evaluated on RefCOCO with IoU, and compare directly with most existing work.
  • Researchers choosing datasets should expect their results to be read against the RefCOCO family and Flickr30K, since those benchmarks dominate the literature and the model-dataset overlap analysis.
  • The reported underuse of remote sensing, dialogue, and construction safety datasets implies that domain-specific grounding is a relatively open space where new benchmarks and models would face less entrenched competition.
  • If the field adopts the survey's recommendations for harmonized metrics and cross-domain benchmarks, reported performance across papers becomes more comparable and industrial adoption becomes more feasible.
  • The paper's claim to be the first dedicated survey would make it the entry point that later surveys must position themselves against.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The frequency figures describe a corpus, not the truth of the field: any update to the search date or screening rules could shift the percentages, so the specific numbers should be treated as provisional census results rather than fixed properties of transformer grounding.
  • The paper's observation that textual encoding still often uses recurrent models in the surveyed studies may be an artifact of corpus selection and publication lag; a search extended through 2025 and 2026 would likely find LLM-based text encoders replacing earlier RNN usage.
  • A natural test of the industrial-readiness claim would be to run the identified dominant models on the underused domain-specific datasets; if performance degrades sharply, the survey's proposed cross-domain benchmark unification would be the highest-value next step.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

5 major / 5 minor

Summary. This manuscript claims to be the first comprehensive systematic literature review of transformer-based spatial grounding in images, covering 2018 to 2025. It reports a systematic methodology (PICO search, Tollgate screening, quality assessment, data extraction) and presents quantitative trends on datasets, model architectures, evaluation metrics, and industrial applicability. The central claim is that the reported frequency distributions accurately describe how the field builds and evaluates spatial grounding models.

Significance. If the quantitative synthesis were properly supported, this survey could serve as a useful community resource: it addresses a real gap, formulates explicit research questions, and attempts a reproducible SLR workflow with inclusion/exclusion criteria and a quality checklist. However, the central statistical claims are currently not auditable. The manuscript contains direct contradictions in the reported corpus size (60 vs. 45 vs. 46), no included-study list is provided, and several figures, tables, and formulas are internally inconsistent. Because the paper's value is its data-driven synthesis, these issues are load-bearing rather than cosmetic. The qualitative observations may be broadly plausible, but the headline percentages (e.g., IoU in 65% of studies, TransVG in 13%, RefCOCO in 57%) cannot yet be verified.

major comments (5)
  1. [§3.1.3, §7.4, §10] The corpus size is internally contradictory: Section 3.1.3 states that the Tollgate process identified 60 primary studies, while Table 2 sums to 45 and the following paragraph says 'From the 45 primary studies selected'; Section 7.4 refers to 'the 46 reviewed papers'; Section 10 again says '45 peer-reviewed studies'. No included-study list or per-study extraction record is provided, so the reader cannot determine which count is correct or reconstruct the corpus. Every frequency reported in Sections 7.2–7.4 depends on this corpus, so these contradictions must be reconciled and the included studies must be listed.
  2. [§7.2, §7.4] The reported percentages are inconsistent with any single corpus size. In §7.2, RefCOCO has frequency 24 and is described as 57%, RefCOCO+ and RefCOCOg have frequency 22 and are described as 52%, and Flicker30K has frequency 18 and is described as 43%; these percentages imply a denominator near 42, not 45 or 46. In §7.4, IoU (≈65%), Accuracy (43%), Recall (28%), Precision (22%), and F1-score (13%) similarly need to be checked against a documented denominator and against the counts in Table 9. The authors should recompute all percentages from a single, explicitly stated corpus size and make the per-study metric coding available.
  3. [§5.2, §5.3] Sections 5.2 and 5.3 make contradictory claims about the dominant methodological pattern. Section 5.2 states that 'a clear methodological pattern emerges where authors predominantly utilize modern transformer architectures for vision tasks, while relying on more traditional sequential models such as RNNs and LSTMs for textual processing.' Section 5.3 states that 'For visual data, most authors have relied on well-established CNNs' and that 'textual modalities ... are modeled predominantly using modern transformer-based language models.' As written, both sections claim to describe the surveyed literature, so they cannot both be true. The authors need to clarify whether these are two distinct subpatterns, report the number of studies supporting each, or remove the 'predominantly/most' language.
  4. [§7.2.1, §7.2.2, §7.5, Figures 5–7] Several figure and table inconsistencies prevent verification of the presented results. Figures 5, 6, and 7 all carry the caption 'Identification of Dominant Datasets in Terms of Instances' even though the surrounding text says they show dataset sizes, ROI formats, and model usage frequencies, respectively. Section 7.2.2 contains the placeholder 'As shown in Figure Y'. Section 7.2.1 states that VDC-UGC is the largest dataset with over 14 million images, but VDC-UGC does not appear in Table 6. Section 7.5 lists GTD-HWT as a dataset, but GTD-HWT is a model in Table 7. These issues must be corrected and the figures relabeled.
  5. [§6.2] The formulas in Section 6.2 are garbled and cannot be used as stated. For example, the METEOR formula is rendered as 'METEOR = (1 − Penalty) × 10 P R R + 9 P' with 'where PP denotes precision and RR recall'; the NDCG formula is 'NDCG@k = DCG@k IDCG@k' with no operator; the precision section reads 'Where TPTPTP and FPFPFP are true positives and false positives respectively'; and the SPICE discussion contains the fragment 'relations matter object.' Additionally, the metric is alternately called 'IDEr' and 'CIDEr' in the same subsection. The authors should provide clean, correctly typeset formulas and consistent nomenclature, since the section explicitly claims to document these metrics.
minor comments (5)
  1. [Throughout] The manuscript contains numerous typos and misspellings, including 'Publicaiton' (§3.1.3), 'Qustion' (§4.2), 'Tranformers' (§5.2 and §8), 'V ision Tranformers' (§2), 'Datasests' (§4.4), and inconsistent 'Flicker30K' vs. 'Flickr30K' spellings. A thorough proofreading pass is needed.
  2. [§7.2.2] The sentence 'As shown in Figure Y' should be replaced with the actual figure reference once the figure captions are corrected.
  3. [§5.2] Section 5.2 says examples include 'MDETR [8] [13] [33] [13]' with [13] duplicated; the duplicate citation should be removed.
  4. [Table 6] In Table 6, the MSCOCO row cites references including [4], and the Flicker30K row cites many references, but the mapping between studies and datasets is not explained anywhere; the authors should clarify how dataset frequencies were derived from the primary studies.
  5. [§2] The claim of being 'the first comprehensive survey dedicated to evaluating transformer-based models specifically for spatial grounding in images' is asserted but not critically compared with the cited survey in [15] ('Towards visual grounding: A survey'); the authors should state concretely what differentiates their coverage from that prior work.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the survey's synthesis is a corpus-frequency exercise, not a derivation, and no claimed result reduces to its own selection criteria.

full rationale

This manuscript is a systematic literature review, so the usual circularity patterns (fitted inputs renamed as predictions, self-citation chains, ansatz smuggled in via citation, uniqueness imported from authors) do not apply. The claimed contribution is a quantitative synthesis of transformer-based spatial grounding research from 2018 to 2025: reported findings such as TransVG appearing in ~13% of surveyed papers, RefCOCO in 57%, and IoU in ~65% are frequency counts computed over the corpus assembled in Section 3. The corpus is selected using a Boolean query and inclusion criteria that require transformer-based architectures and grounding-related tasks, but this does not force any particular answer to the research questions about which specific transformer architectures, datasets, or metrics are prevalent; a different set of transformer papers could have produced different frequencies. The 'first comprehensive survey' claim is a novelty assertion, not a derivational result. The paper's internal inconsistencies (60 vs. 45 vs. 46 studies, no included-study list) and corpus representativeness are reproducibility and validity concerns, not circularity: even if the corpus is biased or incomplete, the reasoning from corpus to frequencies is not self-referential. No equation in the paper defines an output in terms of the quantity it is supposed to predict, and no load-bearing argument relies on a self-citation. Therefore no circular step is present; the appropriate finding is no significant circularity.

Assumptions & free parameters 0 free parameters · 3 assumptions · 0 invented entities

The review rests on three unverified assumptions: corpus completeness, accurate transcription of dataset attributes, and the validity of borrowed quality criteria. It introduces no new entities or fitted parameters. The main burden is that the corpus and extraction are not auditable from the paper alone, so the load-bearing assumption of a complete and representative corpus cannot be checked.

assumptions (3)
  • domain assumption The Boolean query and Tollgate screening yield a complete and representative corpus of transformer-based spatial grounding studies.
    All quantitative conclusions in Section 7 depend on this; no included-study list or inter-rater reliability is provided.
  • domain assumption The dataset attributes reported in Table 6, including sizes, annotation types, and ROI formats, are accurate transcriptions from the cited sources.
    The statistical trends are built on this table, but no extraction data or source-level verification is provided in the paper.
  • domain assumption The quality checklist and A/B/C tiering criteria, including JIF greater than or equal to 1.5 and at least 50 survey respondents, are valid for selecting computer vision papers.
    These criteria were borrowed from software-engineering systematic reviews and are applied without validation to visual grounding model papers, as described in Section 3.1.5.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Transformer-based Spatial Grounding: A Comprehensive Survey." pith.science (2026). https://pith.science/paper/G4D34MAT

@misc{pith2026250712739,
  author       = {Pith},
  title        = {Pith review of: Transformer-based Spatial Grounding: A Comprehensive Survey},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/G4D34MAT}},
  note         = {Machine review of arXiv:2507.12739}
}
read the original abstract

Spatial grounding, the process of associating natural language expressions with corresponding image regions, has rapidly advanced due to the introduction of transformer-based models, significantly enhancing multimodal representation and cross-modal alignment. Despite this progress, the field lacks a comprehensive synthesis of current methodologies, dataset usage, evaluation metrics, and industrial applicability. This paper presents a systematic literature review of transformer-based spatial grounding approaches from 2018 to 2025. Our analysis identifies dominant model architectures, prevalent datasets, and widely adopted evaluation metrics, alongside highlighting key methodological trends and best practices. This study provides essential insights and structured guidance for researchers and practitioners, facilitating the development of robust, reliable, and industry-ready transformer-based spatial grounding models.

Figures

Figures reproduced from arXiv: 2507.12739 by the authors.

Figure 1
Figure 1. This methodology enables us to identify current trends, common practices, research gaps, and future directions [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Publications Screening Process (Tollgate Approach) [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Transformer Based Spatial Grounding Publications per Year and Year-Over-Year Growth [PITH_FULL_IMAGE:figures/full_fig_p018_3.png] view at source ↗
Figures from the paper (9 more)
Figure 4
Figure 4. Figure 4: Dominant Datasets in Transformer Based Spatial Grounding [PITH_FULL_IMAGE:figures/full_fig_p018_4.png]
Figure 5
Figure 5. Figure 5: Identification of Dominant Datasets in Terms of Instances [PITH_FULL_IMAGE:figures/full_fig_p019_5.png]
Figure 6
Figure 6. Figure 6: Identification of Dominant Datasets in Terms of Instances [PITH_FULL_IMAGE:figures/full_fig_p020_6.png]
Figure 7
Figure 7. Figure 7: Identification of Dominant Datasets in Terms of Instances [PITH_FULL_IMAGE:figures/full_fig_p020_7.png]
Figure 8
Figure 8. Figure 8: Transformer Models Vs. Datasets 7.4 Evaluation Metrics Analysis We conducted a comprehensive analysis of evaluation metrics utilized in transformer-based spatial grounding studies, summarized in [PITH_FULL_IMAGE:figures/full_fig_p021_8.png]
Figure 9
Figure 9. Figure 9: Evaluation Metrics Analysis [PITH_FULL_IMAGE:figures/full_fig_p022_9.png]
Figure 10
Figure 10. Figure 10: Comparative Heatmap into the readiness and utility of existing models for industrial applications. The findings are structured across five subsections, supported by quantitative evidence from curated tables and visualizations. 8.1 Industrial Applicaiton Domain Mapping…
Figure 11
Figure 11. Figure 11: Distribution of Transformer Models across Industrial Domains [PITH_FULL_IMAGE:figures/full_fig_p023_11.png]
Figure 12
Figure 12. Figure 12: Per-Domain Precision, Recall, and F1-Score Comparison [PITH_FULL_IMAGE:figures/full_fig_p024_12.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

44 extracted references · 43 canonical work pages

  1. [13]

    W. Su, X. Zhu, Y . Cao, B. Li, L. Lu, F. Wei, J. Dai, Vl-bert: Pre-training of generic visual-linguistic representations, DOI:10.48550/arXiv.1908.08530 (2019)

    arXiv doi:10.48550/arxiv.1908.08530 2019

  2. [1]

    Jiang, Q

    M. Jiang, Q. Huang, L. Zhang, X. Wang, P. Zhang, Z. Gan, J. Diesner, J. Gao, Tiger: Text-to-image grounding for image caption evaluation, EMNLP-IJCNLP (2019). doi:10.18653/v1/D19-1220

    work page doi:10.18653/v1/d19-1220 2019

  3. [2]

    J. Lu, D. Batra, D. Parikh, S. Lee, Vilbert: Pretraining task-agnostic visiolinguistic representations for vision-and- language tasks, Advances in neural information processing systems 32 (2019)

    work page 2019

  4. [3]

    Salgotra, P

    G. Salgotra, P. Abrol, A. Selwal, A survey on automatic image captioning approaches: Contemporary trends and future perspectives, Archives of Computational Methods in Engineering (2024) 1–39

    work page 2024

  5. [4]

    S. Liu, Z. Zeng, T. Ren, F. Li, H. Zhang, J. Yang, Q. Jiang, C. Li, J. Yang, H. Su, et al., Grounding dino: Marrying dino with grounded pre-training for open-set object detection, in: European Conference on Computer Vision, Springer, 2024, pp. 38–55

    work page 2024

  6. [5]

    S. Chen, B. Li, Multi-modal dynamic graph transformer for visual grounding, in: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022, pp. 15534–15543

    work page 2022

  7. [6]

    Silberer, M

    C. Silberer, M. Pinkal, Grounding semantic roles in images, in: Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing, 2018, pp. 2616–2626

    work page 2018

  8. [7]

    J. Y . Koh, R. Salakhutdinov, D. Fried, Grounding language models to images for multimodal inputs and outputs, in: International Conference on Machine Learning, PMLR, 2023, pp. 17283–17300

    work page 2023

Show all 44 references
  1. [8]

    A. Jain, N. Gkanatsios, I. Mediratta, K. Fragkiadaki, Bottom up top down detection transformers for language grounding in images and point clouds, in: European Conference on Computer Vision, Springer, 2022, pp. 417–433

    2022

  2. [9]

    R. A. Yeh, M. N. Do, A. G. Schwing, Unsupervised textual grounding: Linking words to image concepts, in: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2018, pp. 6125–6134

    2018

  3. [10]

    Kiros, W

    J. Kiros, W. Chan, G. Hinton, Illustrative language understanding: Large-scale visual grounding with image search, in: Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (V olume 1: Long Papers), 2018, pp. 922–933

    2018

  4. [11]

    K. Li, J. Li, D. Guo, X. Yang, M. Wang, Transformer-based visual grounding with cross-modality interaction, ACM Transactions on Multimedia Computing, Communications and Applications 19 (6) (2023) 1–19

    2023

  5. [12]

    T. Li, C. Wang, S. Tian, B. Zhang, F. Wu, Y . Tang, H. Zhang, Tacmt: Text-aware cross-modal transformer for visual grounding on high-resolution sar images, ISPRS Journal of Photogrammetry and Remote Sensing 222 (2025) 152–166

    2025

  6. [14]

    J. Deng, Z. Yang, T. Chen, W. Zhou, H. Li, Transvg: End-to-end visual grounding with transformers, in: Proceedings of the IEEE/CVF International Conference on Computer Vision, 2021, pp. 1769–1779

    2021

  7. [15]

    L. Xiao, X. Yang, X. Lan, Y . Wang, C. Xu, Towards visual grounding: A survey, Computer Vision (2024)

    2024

  8. [16]

    L. A. Hendricks, R. Hu, T. Darrell, Z. Akata, Grounding visual explanations, in: Proceedings of the European conference on computer vision (ECCV), 2018, pp. 264–279

    2018

  9. [17]

    Kitchenham, O

    B. Kitchenham, O. P. Brereton, D. Budgen, M. Turner, J. Bailey, S. Linkman, Systematic literature reviews in software engineering–a systematic literature review, Information and software technology 51 (1) (2009) 7–15

    2009

  10. [18]

    Afzal, R

    W. Afzal, R. Torkar, R. Feldt, A systematic review of search-based testing for non-functional system properties, Information and Software Technology 51 (6) (2009) 957–976

    2009

  11. [19]

    Sharma, D

    H. Sharma, D. Padha, Domain-specific image captioning: a comprehensive review, International Journal of Multimedia Information Retrieval 13 (2) (2024) 20

    2024

  12. [20]

    H. Zhao, J. T. Zhou, Y .-S. Ong, Word2pix: Word to pixel cross-attention transformer in visual grounding, IEEE Transactions on Neural Networks and Learning Systems 35 (2) (2022) 1523–1533

    2022

  13. [21]

    J. Deng, Z. Yang, D. Liu, T. Chen, W. Zhou, Y . Zhang, H. Li, W. Ouyang, Transvg++: End-to-end visual grounding with language conditioned vision transformer, IEEE transactions on pattern analysis and machine intelligence 45 (11) (2023) 13636–13652

    2023

  14. [22]

    F. Chen, F. Meng, X. Chen, P. Li, J. Zhou, Multimodal incremental transformer with visual grounding for visual dialogue generation, arXiv preprint arXiv:2109.08478 (2021)

    2021 arXiv

  15. [23]

    Y . Zhan, Z. Xiong, Y . Yuan, Rsvg: Exploring data and models for visual grounding on remote sensing data, IEEE Transactions on Geoscience and Remote Sensing 61 (2023) 1–13. 26

    2023

  16. [24]

    Y . Zhao, Y . Chen, R. Yao, S. Xiong, X. Lu, Context-driven and sparse decoding for remote sensing visual grounding, Information Fusion (2025) 103296

    2025

  17. [25]

    K. Li, D. Wang, H. Xu, H. Zhong, C. Wang, Language-guided progressive attention for visual grounding in remote sensing images, IEEE Transactions on Geoscience and Remote Sensing (2024)

    2024

  18. [26]

    R. Yao, S. Xiong, Y . Zhao, Y . Rong, Visual grounding with multi-modal conditional adaptation, in: Proceedings of the 32nd ACM International Conference on Multimedia, 2024, pp. 3877–3886

    2024

  19. [27]

    C. Chen, S. Yang, H. Wu, L. Liao, Z. Zhang, A. Wang, W. Sun, Q. Yan, W. Lin, Q-ground: Image quality grounding with large multi-modality models, in: Proceedings of the 32nd ACM International Conference on Multimedia, 2024, pp. 486–495

    2024

  20. [28]

    Rasheed, M

    H. Rasheed, M. Maaz, S. Shaji, A. Shaker, S. Khan, H. Cholakkal, R. M. Anwer, E. Xing, M.-H. Yang, F. S. Khan, Glamm: Pixel grounding large multimodal model, in: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024, pp. 13009–13018

    2024

  21. [29]

    J. Liu, W. Fang, P. E. Love, T. Hartmann, H. Luo, L. Wang, Detection and location of unsafe behaviour in digital images: A visual grounding approach, Advanced Engineering Informatics 53 (2022) 101688

    2022

  22. [30]

    Y . Chen, L. Zhan, Y . Zhao, S. Xiong, X. Lu, Vgrss: Datasets and models for visual grounding in remote sensing ship images, IEEE Transactions on Geoscience and Remote Sensing (2025)

    2025

  23. [31]

    Y . Li, H. Liu, Q. Wu, F. Mu, J. Yang, J. Gao, C. Li, Y . J. Lee, Gligen: Open-set grounded text-to-image generation, in: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2023, pp. 22511–22521

    2023

  24. [32]

    L. J. Nixon, Do deep learning models accurately measure visual destination image? a comparison of a fine-tuned model to past work, Information Technology & Tourism 26 (3) (2024) 377–406

    2024

  25. [33]

    M. Qu, Y . Wu, W. Liu, Q. Gong, X. Liang, O. Russakovsky, Y . Zhao, Y . Wei, Siri: A simple selective retraining mechanism for transformer-based visual grounding, in: European Conference on Computer Vision, Springer, 2022, pp. 546–562

    2022

  26. [34]

    H. Yao, L. Wang, C. Cai, W. Wang, Z. Zhang, X. Shang, Language conditioned multi-scale visual attention networks for visual grounding, Image and Vision Computing 150 (2024) 105242

    2024

  27. [35]

    Y . Ding, D. Wang, K. Li, X. Zhao, Y . Wang, Visual grounding of remote sensing images with multi-dimensional semantic-guidance, Pattern Recognition Letters (2025)

    2025

  28. [36]

    L. Li, Y . Cai, J. Wang, C. Wu, Q. Huang, Q. Li, Decoding before aligning: Scale-adaptive early-decoding transformer for visual grounding, Neurocomputing 633 (2025) 129756

    2025

  29. [37]

    J. Wu, C. Wu, Y . Wei, Q. Xu, F. Gong, Learning contrastive semantic decomposition for visual grounding, Neural Networks (2025) 107593

    2025

  30. [38]

    M. Li, L. Sigal, Referring transformer: A one-step approach to multi-task visual grounding, Advances in neural information processing systems 34 (2021) 19652–19664

    2021

  31. [39]

    R. Shao, T. Wu, Z. Liu, Detecting and grounding multi-modal media manipulation, in: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023, pp. 6904–6913

    2023

  32. [40]

    L. Li, Z. Zheng, Y . Bu, C. Wu, S. Huang, Q. Huang, Y . Cai, Grouped top-down reasoning with hierarchical window transformer for visual grounding, Information Processing & Management 62 (6) (2025) 104222

    2025

  33. [41]

    J. Wu, C. Wu, F. Wang, L. Wang, Y . Wei, Improving visual grounding with multi-scale discrepancy information and centralized-transformer, Expert Systems with Applications 247 (2024) 123223

    2024

  34. [42]

    J. Cho, Y . Yoon, S. Kwak, Collaborative transformers for grounded situation recognition, in: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022, pp. 19659–19668

    2022

  35. [43]

    J. Li, Z. Mao, H. Li, W. Chen, Y . Zhang, Exploring visual relationships via transformer-based graphs for enhanced image captioning, ACM Transactions on Multimedia Computing, Communications and Applications 20 (5) (2024) 1–23

    2024

  36. [44]

    Belhadi, Y

    A. Belhadi, Y . Djenouri, A. N. Belbachir, T. Michalak, G. Srivastava, Shapley visual transformers for image-to-text generation, Applied Soft Computing 166 (2024) 112205. 27

    2024

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.