Pith. sign in

REVIEW 2 major objections 4 minor 41 references

ReCap: Event-Aware Image Captioning with Article Retrieval and Semantic Gaussian Normalization

T0 review · 2 major / 4 minor · reviewed 2026-08-05 · deepseek-v4-flash

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

Pith's one-line read Event-aware captioning is a retrieval problem first: retrieve the right article, fuse its context, then normalize caption length against CIDEr's penalty.

desk verdict Retrieval engineering is solid and honestly reported, but the captioning gain is mostly CIDEr tuning via a length-penalty module, not a real caption-quality improvement. read the letter →

arxiv 2509.01259 v1 pith:4CNNAYFB submitted 2025-09-01 cs.CV

Thinh-Phuc Nguyen , Thanh-Hai Nguyen , Gia-Huy Dinh , Lam-Huy Nguyen , Minh-Triet Tran , Trung-Nghia Le This is my paper
classification cs.CV
keywords imagecaptioningevent-enrichedretrievalDINOv2mutualnearestneighborrerankinglargelanguagemodelCIDErnewsimages
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

ReCap claims that event-enriched image captioning is best treated as a retrieval problem first, a language-fusion problem second, and a metric-alignment problem last. Given a heavily transformed news photo, the pipeline retrieves the matching article using DINOv2 global features and patch-level mutual-nearest-neighbor reranking, then asks a large language model to fuse the generic visual caption, the article's web caption, and an article summary into a single enriched caption. A final Semantic Gaussian Normalizer truncates or pads the caption to around 104–105 words to soften the Gaussian length penalty in CIDEr. Reporting the second-highest overall score of 0.54666 on the private test set of the EVENTA 2025 Track 1 challenge, the paper argues that this recipe works: correct retrieval enables grounded context, and the normalizer converts that context into metric-visible gains. If right, the approach gives newsrooms and archives a practical template for captions that say what event is happening, not just what is in the frame.

What carries the argument

The load-bearing object is the Semantic Gaussian Normalizer, a post-processing module that truncates captions longer than roughly 104–105 words, removes low-value words before named entities, and appends named entities to short captions; it exists to counteract the Gaussian length penalty in the CIDEr metric, which exponentially discounts captions whose word count differs from a reference. The retrieval side is carried by two-stage DINOv2 matching: global cosine similarity selects top-100 candidates, then patch-level mutual nearest neighbor similarity (Equation 1) reranks them by bidirectional per-patch maximum cosine similarity.

What would settle it

Take a held-out sample of query images from the same event distribution with several reference captions each. Run ReCap with and without Semantic Gaussian Normalization, then compare CIDEr against human raters who choose which caption better identifies the event and its participants. If the truncated or padded captions score higher on CIDEr but not on human relevance, the claim that normalization improves caption quality fails; likewise if varying the threshold by a few words on the private distribution changes CIDEr by more than the reported gain.

Watch

Extended reading notes

Core claim

The paper's central claim is that event-enriched image captioning can be broken into three manageable stages, and that each stage contributes measurable gains. The first stage is a two-stage article retrieval system: DINOv2 global embeddings select the top-100 candidates by cosine similarity, then patch-level mutual maximum nearest neighbor similarity reranks them, lifting Recall@1 from 0.921 to 0.977 on the private test set. The second stage extracts three context signals from the retrieved article—a generic visual caption produced by Qwen2.5-VL, a web caption matched by CLIP, and an article summary generated by Qwen3—and prompts Qwen3 to fuse them into one enriched caption, achieving a CLI

Load-bearing premise

The claim that truncating or padding captions to roughly 104–105 words improves caption quality depends on that threshold, picked by repeated submissions to the public test set, still being optimal on the private test set and on the CIDEr gains reflecting genuine semantic enrichment rather than metric fitting.

Editorial extensions

If this is right

  • Correct article retrieval becomes the bottleneck for event captions: ReCap's R@1 rises from 0.921 with global features alone to 0.977 after patch-level mutual-nearest-neighbor reranking, and captioning quality depends on retrieving the right article.
  • CIDEr can be improved substantially by post-processing alone: adding the Semantic Gaussian Normalizer raises CIDEr from 0.145 to 0.205 on the private test set without any retraining.
  • Fusing a generic visual caption, a web caption, and an article summary through an LLM produces captions that score high on CLIP alignment (0.870), suggesting article context can be bound to the visible image through prompting.
  • The full pipeline is competitive as a deployable system, ranking 2nd on the challenge leaderboard with an overall score of 0.54666.

Reading between the lines

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

  • Inference: Part of the reported CIDEr gain likely reflects length calibration rather than semantic enrichment, because CIDEr's Gaussian term penalizes any length mismatch and the 104–105 word threshold was tuned to the training-set length distribution; a fair comparison would control for caption length.
  • Inference: The retrieval-plus-rerank recipe should transfer to other news-image retrieval tasks where queries are heavily transformed, since the gain comes from patch-level correspondence rather than task-specific tuning.
  • Inference: A testable extension is to replace the exhaustive patch-level mutual nearest neighbor matching with a learned or quantized patch retriever; if R@1 stays near 0.977 at lower cost, the reranking stage becomes practical at much larger scale.
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

2 major / 4 minor

Summary. The paper introduces ReCap, a pipeline for the EVENTA 2025 Track 1 event-enriched image retrieval and captioning challenge. Retrieval uses DINOv2 global descriptors for top-100 candidate selection followed by patch-level mutual-nearest-neighbor similarity re-ranking. Captioning extracts a generic Qwen2.5-VL caption, a crawled web caption, and an article summary, and fuses them with Qwen3 under an instruction prompt. A Semantic Gaussian Normalizer (truncation with or without entity preservation, plus entity enrichment) is applied to align caption length with the CIDEr Gaussian penalty. On the private test set the system reports overall 0.54666 (2nd place), with strong retrieval (mAP 0.982, R@1 0.977, R@10 0.988, CLIPScore 0.870) but CIDEr only 0.205. Ablations show retrieval gains from re-ranking and monotone CIDEr/overall gains from the normalizer.

Significance. The retrieval component is a solid engineering contribution: the two-stage DINOv2 global+patch reranking is simple, reproducible, and yields large R@1 gains (0.921→0.977 on private, Table 2). The authors release code and are transparent about their public-test hyperparameter search. However, the central captioning claim is currently unproven: the CIDEr improvement from the Semantic Gaussian Normalizer is derived from the metric's length penalty and rare-word weighting, and the length threshold was selected by probing the public test set. No human evaluation, semantic metrics, or validation split is provided. If the paper is intended as a challenge-system report, the metric-alignment contribution can be stated as such; as a captioning-quality claim, it needs additional evidence.

major comments (2)
  1. [Sec. 4.3.1, Table 2] The paper's central captioning result is the private-test CIDEr improvement from 0.145 to 0.205 attributable to Semantic Gaussian Normalizer. However, the operating point (104-105 word threshold) was selected by repeated submissions to the public test set, as the authors state in Sec. 4.3.1. No held-out validation split was used. Under these conditions the private-test improvement is not an independent estimate of the method's effect; it is the result of test-set model selection. Please report a validation-split threshold search, or show that the private-test score is stable over a range of thresholds (e.g., 103-108 words), before claiming that SGN generalizes.
  2. [Sec. 3.4, Eq. (2)] The SGN design is derived from the CIDEr formula: Eq. (2) contains the Gaussian length penalty exp(-(l(ci)-l(sij))^2/(2σ^2)) and TF-IDF weighting that rewards rare n-grams. Truncating to a pre-chosen length and appending named entities therefore raises CIDEr by construction. Table 2 shows that the entire improvement of SGN is confined to CIDEr (0.145→0.205); CLIPScore is unchanged (0.870). The claim that SGN 'enhances fluency and relevance' and produces 'semantically normalized' captions is not supported by any human evaluation or auxiliary semantic metric. Add human ratings, entity-precision/recall, or factual-consistency evaluation, or revise the claim to state that SGN is a metric-alignment post-processor.
minor comments (4)
  1. [Sec. 3.3] Typo: 'employst' should be 'employs'. Also 'we employt' appears in the caption-generation paragraph.
  2. [Sec. 3.2 vs. Sec. 4.1] Section 3.2 says DINOv2 ViT-g/14 with registers, but Section 4.1 says DINOv2-Large (1B parameters). Clarify which backbone was used for the reported results.
  3. [Table 1 vs. Sec. 4.3.1] Table 1 lists Qwen3 temperature 0.6, top-p 0.95, top-k 20, while Section 4.3.1 says temperature 0.85, top-p 0.8, top-k 50 was suboptimal. Please state explicitly which configuration produced the final leaderboard numbers.
  4. [Figure 6] The caption-length distribution plot lacks axis labels and a legend, making it hard to verify the claimed 90-120 word operating range. Add axis labels and describe the data split.

Circularity Check

2 steps flagged · score 6.0 of 10

SGN's CIDEr gains are a direct implementation of CIDEr-D's length and rarity terms, and the truncation threshold is tuned on the public test set; retrieval results are not circular.

  1. self definitional [Eq. (2), Sec. 3.4 (Semantic Gaussian Normalizer), Table 2]
    "This equation has the Gaussian penalty as a heavy penalty for captions which have even minor differences in caption length as compared to the reference caption. This penalty is indicated by the term e^{-(l(c_i)-l(s_ij))^2/2\sigma^2} which fines exponentially the caption length difference between the generated and the reference. ... For such motivation, we propose a trade-off strategy to manage caption length. If a generated caption exceeds a predetermined threshold, we truncate it. Conversely, if its length falls below this threshold, we enrich the caption with additional words. The primary ob"

    SGN is constructed directly from CIDEr-D's Eq. (2): Gaussian/Semantic Normalizers truncate captions so the exponential length penalty term e^{-...} is driven toward 1, and Entity Enricher appends named entities because the paper states CIDEr assigns higher scores to rare words. The paper then presents the resulting CIDEr jump in Table 2 (0.145 to 0.205) as the module's contribution to 'fluency and relevance.' Since the module's operations are exactly the terms in CIDEr-D, the CIDEr improvement is by construction rather than independent evidence of caption quality. CLIPScore stays 0.870 before and after SGN, so no separate caption-quality metric moves.

  2. fitted input called prediction [Sec. 4.3.1 (Hyperparameter Search Using Public Test), Table 2]
    "During the development phase, we performed an extensive hyperparameter search on the test set, specifically focusing on the maximum word threshold for our truncation strategies. ... 108 words: Achieved an overall score of 0.54. ... 104 words: Achieved an overall score of 0.55. ... These results indicate that a maximum word threshold of 104-105 words yielded the best performance on the test set for our algorithm."

    The truncation threshold is fitted by repeated submissions to the public test set. This fitted threshold is then used in SGN, whose private-test CIDEr gain (0.145 to 0.205) is reported in Table 2. Thus the captioning gain attributed to SGN is the outcome of tuning the module's key hyperparameter on the test distribution, not a held-out validation of a pre-specified method. The paper itself calls the result 'a highly optimal operating point,' confirming the gain is a metric fit rather than a parameter-free prediction.

full rationale

The retrieval components are not circular: DINOv2 global retrieval plus patch-level mutual-nearest-neighbor reranking is evaluated directly on private-test mAP/R@1/R@10, with no fitted parameter renamed as a prediction. The self-citations [14] and [27] only supply the benchmark/task and are not load-bearing for the method. The captioning contribution, however, is partially circular. SGN explicitly targets CIDEr-D's Gaussian length penalty and rare-word TF-IDF weighting (Sec. 3.4 vs Eq. 2), and its operating threshold is tuned on the public test set via multiple submissions (Sec. 4.3.1). The CIDEr gain attributed to SGN in Table 2 is therefore substantially a by-construction metric fit: CLIPScore is unchanged, so no independent caption-quality signal supports the 'fluency and relevance' claim. Since the private test is genuinely separate and the leaderboard result is external, the circularity is partial rather than total; score 6 reflects that the central captioning result reduces to metric fitting while the retrieval result remains independent.

Assumptions & free parameters 4 free parameters · 5 assumptions · 0 invented entities

The central score rests on benchmark assumptions: correct article exists, visual features are robust, LLM fusion works, and CIDEr is the right reward. The normalizer's threshold is a tuned hyperparameter, and the decoding settings are tuned as well. No new physical or conceptual entities are introduced.

free parameters (4)
  • Caption truncation threshold = 104-105 words
    Selected by searching the public test set via multiple challenge submissions (Sec 4.3.1); peak of training distribution is 90-120 words but exact value tuned on test.
  • Qwen3 decoding temperature/top-p/top-k = T=0.6, top-p=0.95, top-k=20
    Table 1 reports tuned decoding settings; Sec 4.3.1 notes T=0.85, top-p=0.8, top-k=50 gave lower overall score.
  • Number of initial retrieval candidates = 100
    Top-100 candidates in stage 1; chosen by hand, not justified by sensitivity analysis.
  • Entity Enricher target minimum length = not specified
    Paper says a 'desired minimum length' is used to pad under-length captions but does not report the exact value.
assumptions (5)
  • domain assumption The correct article for every query image is present in the article database.
    Retrieval mAP/R@1 presume ground-truth article exists among candidates; task from OpenEvents/EVENTA benchmark, not verified in paper.
  • domain assumption DINOv2 global and patch features preserve identity under heavy image transformations.
    Sec 3.2 relies on this; no per-transform robustness analysis is reported.
  • domain assumption Qwen3 can fuse the generic caption, web caption, and article summary into a factually grounded enriched caption.
    Sec 3.3-4.1 use reasoning mode and prompting; no factual accuracy evaluation is included.
  • domain assumption CLIP image-text matching correctly assigns the scraped web caption to the retrieved image.
    Sec 3.3 Context Extraction uses CLIP matching; matching accuracy is not reported.
  • domain assumption CIDEr-D is an appropriate measure of event-enriched caption quality.
    The method explicitly optimizes Eq. 2; paper provides no evidence that CIDEr correlates with factual or event-level correctness.

how reviews work

0 comments
Cite this review

Pith. "Pith review of ReCap: Event-Aware Image Captioning with Article Retrieval and Semantic Gaussian Normalization." pith.science (2026). https://pith.science/paper/4CNNAYFB

@misc{pith2026250901259,
  author       = {Pith},
  title        = {Pith review of: ReCap: Event-Aware Image Captioning with Article Retrieval and Semantic Gaussian Normalization},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/4CNNAYFB}},
  note         = {Machine review of arXiv:2509.01259}
}
read the original abstract

Image captioning systems often produce generic descriptions that fail to capture event-level semantics which are crucial for applications like news reporting and digital archiving. We present ReCap, a novel pipeline for event-enriched image retrieval and captioning that incorporates broader contextual information from relevant articles to generate narrative-rich, factually grounded captions. Our approach addresses the limitations of standard vision-language models that typically focus on visible content while missing temporal, social, and historical contexts. ReCap comprises three integrated components: (1) a robust two-stage article retrieval system using DINOv2 embeddings with global feature similarity for initial candidate selection followed by patch-level mutual nearest neighbor similarity re-ranking; (2) a context extraction framework that synthesizes information from article summaries, generic captions, and original source metadata; and (3) a large language model-based caption generation system with Semantic Gaussian Normalization to enhance fluency and relevance. Evaluated on the OpenEvents V1 dataset as part of Track 1 in the EVENTA 2025 Grand Challenge, ReCap achieved a strong overall score of 0.54666, ranking 2nd on the private test set. These results highlight ReCap's effectiveness in bridging visual perception with real-world knowledge, offering a practical solution for context-aware image understanding in high-stakes domains. The code is available at https://github.com/Noridom1/EVENTA2025-Event-Enriched-Image-Captioning.

Figures

Figures reproduced from arXiv: 2509.01259 by the authors.

Figure 1
Figure 1. Overall pipeline of the proposed ReCap. A heavily transformed query image is processed by DINOv2 to extract both [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Pipeline for Article and Image Retrieval and Re [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Visualization Patch-level Mutual Nearest Neighbor [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: Caption enriching process: useful information [PITH_FULL_IMAGE:figures/full_fig_p004_4.png]
Figure 5
Figure 5. Figure 5: The graph of the function 𝑓 (𝑥) = 𝑒 −𝑥 2 /(2·𝜎 2 ) , where 𝜎 = 6 is normally used in CIDEr score. The variable 𝑥 indicates the caption length difference between the generated caption and the reference sentence; 𝑓 (𝑥) indicates the penalty factor [PITH_FULL_IMAGE:figur…
Figure 6
Figure 6. Figure 6: Generated caption word length distribution on train [PITH_FULL_IMAGE:figures/full_fig_p005_6.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

41 extracted references · 35 canonical work pages

  1. [1]

    Jyoti Aneja, Aditya Deshpande, and Alexander G Schwing. 2018. Convolutional image captioning. In Proceedings of the IEEE conference on computer vision and pattern recognition. 5561–5570

    work page 2018

  2. [2]

    Ali Furkan Biten, Lluis Gomez, Marçal Rusinol, and Dimosthenis Karatzas. 2019. Good news, everyone! context driven entity-aware captioning for news images. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition . 12466–12475

    work page 2019

  3. [3]

    Kanzhi Cheng, Wenpo Song, Zheng Ma, Wenhao Zhu, Zixuan Zhu, and Jianbing Zhang. 2023. Beyond generic: Enhancing image captioning with real-world knowledge using vision-language pre-training model. In Proceedings of the 31st ACM International Conference on Multimedia . 5038–5047

    work page 2023

  4. [4]

    Kailash A Hambarde and Hugo Proenca. 2023. Information retrieval: recent advances and beyond. IEEE Access 11 (2023), 76581–76604

    work page 2023

  5. [5]

    Simao Herdade, Armin Kappeler, Kofi Boakye, and Joao Soares. 2019. Image captioning: Transforming objects into words. Advances in neural information processing systems 32 (2019)

    work page 2019

  6. [6]

    MD Zakir Hossain, Ferdous Sohel, Mohd Fairuz Shiratuddin, and Hamid Laga

  7. [7]

    Anwen Hu, Shizhe Chen, and Qin Jin. 2020. ICECAP: Information concentrated entity-aware image captioning. In Proceedings of the 28th ACM International Conference on Multimedia. 4217–4225

    work page 2020

  8. [8]

    Ronghang Hu, Amanpreet Singh, Trevor Darrell, and Marcus Rohrbach. 2020. Iterative Answer Prediction with Pointer-Augmented Multimodal Transformers for TextVQA. In CVPR

    work page 2020

Show all 41 references
  1. [9]

    Azhar Jamil, Khalid Mahmood, Monica Gracia Villar, Thomas Prola, Isabel De La Torre Diez, Md Abdus Samad, Imran Ashraf, et al . 2024. Deep learning approaches for image captioning: Opportunities, challenges and future potential. IEEE Access (2024)

    2024

  2. [10]

    Cijo Jose, Théo Moutakanni, Dahyun Kang, Federico Baldassarre, Timothée Darcet, Hu Xu, Daniel Li, Marc Szafraniec, Michaël Ramamonjisoa, Maxime Oquab, et al . 2025. Dinov2 meets text: A unified framework for image-and pixel-level vision-language alignment. In Proceedings of th...

    2025

  3. [11]

    Ting-ting Li, Bo Jiang, Zheng-zheng Tu, Bin Luo, and Jin Tang. 2015. Image matching using mutual k-nearest neighbor graph. In Intelligent Computation in Big Data Era: International Conference of Young Computer Scientists, Engineers and Educators, ICYCSEE 2015, Harbin, China, J...

    2015

  4. [12]

    Fuxiao Liu, Yinghan Wang, Tianlu Wang, and Vicente Ordonez. 2020. Visual news: Benchmark and challenges in news image captioning. arXiv preprint arXiv:2010.03743 (2020)

    2020 arXiv

  5. [13]

    Tianrui Liu, Qi Cai, Changxin Xu, Bo Hong, Jize Xiong, Yuxin Qiao, and Tsung- wei Yang. 2024. Image Captioning in news report scenario. arXiv preprint arXiv:2403.16209 (2024)

    2024 arXiv

  6. [14]

    Nguyen, Minh-Triet Tran, and Trung-Nghia Le

    Hieu Nguyen, Phuc-Tan Nguyen, Thien-Phuc Tran, Minh-Quang Nguyen, Tam V. Nguyen, Minh-Triet Tran, and Trung-Nghia Le. 2025. OpenEvents V1: Large- Scale Benchmark Dataset for Multimodal Event Grounding. InACM International Conference on Multimedia

    2025

  7. [15]

    Sofia Nikiforova, Tejaswini Deoskar, Denis Paperno, and Yoad Winter. 2020. Geo-aware image caption generation. In Proceedings of the 28th International Conference on Computational Linguistics . 3143–3156

    2020

  8. [16]

    Maxime Oquab, Timothée Darcet, Theo Moutakanni, Huy V. Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, Russell Howes, Po-Yao Huang, Hu Xu, Vasu Sharma, Shang-Wen Li, Wojciech Galuba, Mike Rabbat, Mido Assran, Nicolas B...

    2023

  9. [17]

    Jae Sung Park, Chandra Bhagavatula, Roozbeh Mottaghi, Ali Farhadi, and Yejin Choi. 2020. Visualcomet: Reasoning about the dynamic context of a still image. In Computer Vision–ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part V 16 . Springe...

    2020

  10. [18]

    Tingyu Qu, Tinne Tuytelaars, and Marie-Francine Moens. 2023. Visually-aware context modeling for news image captioning. arXiv preprint arXiv:2308.08325 (2023)

    2023 arXiv

  11. [19]

    Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. 2021. Learning transferable visual models from natural language supervision. In International conference on machine learni...

    2021

  12. [20]

    Shihao Shao, Kaifeng Chen, Arjun Karpur, Qinghua Cui, André Araujo, and Bingyi Cao. 2023. Global features are all you need for image retrieval and reranking. In Proceedings of the IEEE/CVF International Conference on Computer Vision. 11036–11046

    2023

  13. [21]

    Palaiahnakote Shivakumara, Anjan Dutta, Trung Quy Phan, Chew Lim Tan, and Umapada Pal. 2011. A novel mutual nearest neighbor based symmetry for text frame classification in video. Pattern Recognition 44, 8 (2011), 1671–1683

    2011

  14. [22]

    Oleg Sidorov, Ronghang Hu, Amanpreet Singh, Marcus Rohrbach, and Trevor Darrell. 2020. TextCaps: A Dataset for Image Captioning with Reading Compre- hension. In ECCV

    2020

  15. [23]

    Fuwen Tan, Jiangbo Yuan, and Vicente Ordonez. 2021. Instance-level image re- trieval using reranking transformers. In proceedings of the IEEE/CVF international conference on computer vision . 12105–12115

    2021

  16. [24]

    Qwen Team. 2025. Qwen2.5-VL. https://qwenlm.github.io/blog/qwen2.5-vl/

    2025

  17. [25]

    Alasdair Tran, Alexander Mathews, and Lexing Xie. 2020. Transform and tell: Entity-aware news image captioning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition . 13035–13045

    2020

  18. [26]

    Duy-Kien Tran, Tien Bui, Khoa Tran, and Minh-Tien Nguyen. 2022. NY- Times800K: A Dataset for Image Captioning with Paragraph-Level Summaries. arXiv preprint arXiv:2201.12321 (2022)

    2022 arXiv

  19. [27]

    Nguyen, Trong- Le Do, Duy-Nam Ly, Viet-Tham Huynh, Khanh-Duy Le, Mai-Khiem Tran, and Trung-Nghia Le

    Thien-Phuc Tran, Minh-Quang Nguyen, Minh-Triet Tran, Tam V. Nguyen, Trong- Le Do, Duy-Nam Ly, Viet-Tham Huynh, Khanh-Duy Le, Mai-Khiem Tran, and Trung-Nghia Le. 2025. Event-Enriched Image Analysis Grand Challenge At ACM Multimedia 2025. In ACM International Conference on Multimedia

    2025

  20. [28]

    Ramakrishna Vedantam, C Lawrence Zitnick, and Devi Parikh. 2015. Cider: Consensus-based image description evaluation. In Proceedings of the IEEE confer- ence on computer vision and pattern recognition . 4566–4575

    2015

  21. [29]

    Oriol Vinyals, Alexander Toshev, Samy Bengio, and Dumitru Erhan. 2016. Show and tell: Lessons learned from the 2015 mscoco image captioning challenge. IEEE transactions on pattern analysis and machine intelligence 39, 4 (2016), 652–663

    2016

  22. [30]

    Ning Xu, Tingting Zhang, Hongshuo Tian, and An-An Liu. 2024. Rule-driven News Captioning. IEEE Transactions on Circuits and Systems for Video Technology (2024)

    2024

  23. [31]

    An Yang, Anfeng Li, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chang Gao, Chengen Huang, Chenxu Lv, et al. 2025. Qwen3 technical report. arXiv preprint arXiv:2505.09388 (2025)

    2025 arXiv

  24. [32]

    Xiaobao Yang, Yang Yang, Sugang Ma, Zhijie Li, Wei Dong, and Marcin Woźniak

  25. [33]

    Quanzeng You, Hailin Jin, Zhaowen Wang, Chen Fang, and Jiebo Luo. 2016. Image captioning with semantic attention. In Proceedings of the IEEE conference on computer vision and pattern recognition . 4651–4659

    2016

  26. [34]

    Zequn Zeng, Yan Xie, Hao Zhang, Chiyu Chen, Bo Chen, and Zhengjue Wang

  27. [35]

    Jingjing Zhang, Shancheng Fang, Zhendong Mao, Zhiwei Zhang, and Yongdong Zhang. 2022. Fine-tuning with multi-modal entity prompts for news image captioning. In Proceedings of the 30th ACM International Conference on Multimedia. 4365–4373

    2022

  28. [36]

    Xuanmeng Zhang, Minyue Jiang, Zhedong Zheng, Xiao Tan, Errui Ding, and Yi Yang. 2020. Understanding image retrieval re-ranking: A graph neural network perspective. arXiv preprint arXiv:2012.07620 (2020)

    2020 arXiv

  29. [37]

    In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition

    Meacap: Memory-augmented zero-shot image captioning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition . 14100–14110

  30. [38]

    Wentian Zhao and Xinxiao Wu. 2023. Boosting entity-aware image captioning with multi-modal knowledge graph. IEEE Transactions on Multimedia 26 (2023), 2659–2670

    2023

  31. [40]

    Sanqiang Zhao, Piyush Sharma, Tomer Levinboim, and Radu Soricut. 2019. Infor- mative image captioning with external sources of information. arXiv preprint arXiv:1906.08876 (2019)

    2019 arXiv

  32. [2019]

    ACM Computing Surveys (CsUR) 51, 6 (2019), 1–36

    A comprehensive survey of deep learning for image captioning. ACM Computing Surveys (CsUR) 51, 6 (2019), 1–36

    2019

  33. [2024]

    Neurocomputing 593 (2024), 127823

    SAMT-generator: A second-attention for image captioning based on multi- stage transformer network. Neurocomputing 593 (2024), 127823

    2024

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.