Pith. sign in

REVIEW 4 major objections 5 minor 33 references

ReFrame: Rectification Framework for Image Explaining Architectures

T0 review · 4 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 ReFrame claims that a Mask R-CNN-based rectification layer, plugged on top of any image-explaining model, corrects hallucinated and missing objects and, on the tested baselines, improves captioning completeness by 81.81% and reduces…

desk verdict ReFrame's headline numbers are an artifact of using Mask R-CNN's detections as both the correction oracle and the evaluation reference; the paper is a useful negative example, not a research result. read the letter →

arxiv 2506.18272 v1 pith:IZ2FWKVG submitted 2025-06-23 cs.CV

Debjyoti Das Adhikary , Aritra Hazra , Partha Pratim Chakrabarti This is my paper · ORCID
classification cs.CV MSC 68T4568T07
keywords ImagecaptioningVisualquestionansweringObjecthallucinationCompletenessInconsistencyMaskR-CNNRectificationframeworkPrompt-basedAI
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

The paper proposes ReFrame, a plug-in layer that sits on top of an existing image-explaining model and corrects the objects that model names. Its claim is that object hallucination and omission can be reduced without retraining the base model, by checking generated object words against the detections of a Mask R-CNN supervisor and replacing disagreeing words with the next most probable accepted token. On the tested baselines, ReFrame improves image-captioning completeness by 81.81% and reduces inconsistency by 37.10%, improves VQA completeness by about 8.9 to 10.6%, and gives a small gain for GPT-4. A sympathetic reader would care because the approach offers a cheap, interpretable way to make multimodal explanations more trustworthy, provided the detector itself is trustworthy.

What carries the argument

The load-bearing mechanism is Mask R-CNN used as an external object supervisor, together with the two metrics that measure success by agreement with that same supervisor. Three adapters connect the supervisor to each domain: a recursive LSTM that expands the caption decoder's search into a tree of candidate captions; a word2vec projection that maps decoder tokens to one of the 80 COCO classes so mismatches can be identified; and a top-k agreement rule for VQA answers plus a re-prompt loop for GPT-4. Both proposed metrics, Inconsistency and Completeness, normalize by the set of objects with rectification, so the detector is simultaneously the correction mechanism and the reference standard.

What would settle it

Take a set of images where human annotators mark objects that Mask R-CNN is known to miss, such as small, occluded, or less common objects, and compare ReFrame's corrected captions with the human lists; if the framework drops or fails to mention those human-visible objects and its completeness score still improves, the assumption that the detector is the ground truth is falsified.

Watch

Extended reading notes

Core claim

At the paper's center is the claim that the two chronic failures of image explanation, hallucinating objects that are absent and omitting objects that are present, can be treated as a post-processing consistency problem. The rectifier, Mask R-CNN, produces an authoritative list of present objects; the base model's output is accepted only where it agrees with that list, and where it disagrees, the next most likely token that the detector recognizes is substituted. For image captioning this is done by generating a tree of candidate captions with a recursive LSTM; for VQA the top-k answers are re-ranked by detector agreement; for GPT-4 the model is re-prompted with the detector's findings. The paper reports that this yields a captioning completeness of 0.540 versus 0.297 at baseline, an inconsistency of 0.200 versus 0.318, and similar or better numbers than the state-of-the-art models it compares against.

Load-bearing premise

The entire correction and the scores both assume that Mask R-CNN's detections are a complete and correct list of the objects in the image, and that word2vec can map every caption word to one of those 80 classes without error.

Editorial extensions

If this is right

  • ReFrame can be mounted on top of existing image captioning and VQA models and improve their object completeness without modifying or retraining the base models.
  • Captioning completeness rises from 0.297 to 0.540 and inconsistency falls from 0.318 to 0.200 on the validation set used.
  • For VQA, the rectified model beats its baseline at top-5, top-10 and top-15 and beats the compared state-of-the-art model on inconsistency at top-10 and top-15.
  • GPT-4's object descriptions become slightly more consistent, with 5.2% lower inconsistency, when re-prompted with the detector's object list.
  • As the paper itself notes, the benefit depends on the base model and the rectifier being trained on the same dataset; outside that setting the rectification ability is limited.

Reading between the lines

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

  • Because the same Mask R-CNN both decides which words to replace and defines the completeness and inconsistency denominators, part of the reported improvement may reflect convergence on the detector's mistakes rather than a genuine reduction in hallucination; a human-annotation comparison would settle this.
  • The framework is intrinsically bounded by the 80 COCO classes and by word2vec's ability to map words to those classes; moving to an open-vocabulary detector is the natural stress test.
  • The same post-hoc supervision pattern could be applied to other generation tasks with a fixed set of checkable entities, such as scene graphs, medical findings, or aerial imagery reports.
  • The prominence-based completeness analysis suggests that detector-based rectification interacts with object size; tuning the detector's confidence or area threshold could trade precision against recall differently in each domain.
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

4 major / 5 minor

Summary. The paper proposes ReFrame, a Mask R-CNN (MRCNN) based rectification layer that is placed on top of three image-explaining architectures: image captioning (Show and Tell), Visual Question Answering (SAN), and prompt-based GPT-4. The rectifier is used to replace or filter predicted object tokens so that they agree with MRCNN's detections, and the authors introduce two object-based metrics, Inconsistency (Eq. 1) and Completeness (Eq. 2), both normalized by the set of objects detected by the rectifier. The paper reports large improvements for captioning (81.81% completeness increase, 37.10% inconsistency reduction), substantial VQA gains, and small GPT-4 gains, and claims to surpass current state-of-the-art models.

Significance. If the claimed improvements were verified against independent object-level ground truth, a model-agnostic rectifier built on a fixed object detector would be a practical contribution to reducing object hallucination and omission in multimodal explanation systems. The paper deserves credit for proposing a pluggable architecture, for attempting object-based precision metrics beyond BLEU/CIDEr, and for explicitly acknowledging in Section 6 that the baseline and rectifier are trained on the same dataset. However, because the proposed evaluation metrics are defined against the same detector that drives the rectification, the headline quantitative results do not currently provide evidence of improved image understanding; independent evaluation is required before the central claim can be assessed.

major comments (4)
  1. [§4.3, Eqs. (1)-(2); §3.1-3.3] The central evaluation is circular. Completeness in Eq. (2) is the fraction of the rectifier's detected object set that is also mentioned by the explanation, and Inconsistency in Eq. (1) is the set difference normalized by the same rectifier set. Sections 3.1-3.3 describe replacing or filtering tokens so that they agree with the rectifier's detections. Consequently, any post-rectification explanation is scored against the same oracle that constructed it, so the reported gains (e.g., 81.81% completeness and 37.10% inconsistency reduction in Section 5.1) are largely tautological. The manuscript needs an evaluation against independent object-level annotations, such as COCO ground-truth object labels or human judgments, before improvements can be attributed to better image understanding.
  2. [Abstract vs. §5.2, §5.3] The numerical claims are internally inconsistent. The abstract reports an average VQA inconsistency reduction of 37.10%, while Section 5.2 reports drops of 61%, 62.73%, and 64.31% for top-5, top-10, and top-15, respectively; the abstract states a GPT-4 completeness gain of 0.01%, while Section 5.3 reports 0.02%. These discrepancies must be reconciled before the results can be considered reliable.
  3. [§3, §4.2.3] The word2vec-based token mapping is not validated. Section 3 states that the entire decoder vocabulary is mapped to the 80 MRCNN classes via GoogleNews word vectors, but no analysis is provided of mapping accuracy or of cases where a valid token is replaced by an unrelated COCO class. Since this mapping is the mechanism by which tokens are corrected, its errors are inherited by every downstream metric.
  4. [§4.2.3, §5.3] The GPT-4 arm is not a reproducible experiment. Section 4.2.3 states that rectification prompts were manually fed to GPT-4, and Section 5.3 reports a 5.2% inconsistency improvement on that basis. Without a systematic protocol specifying the number of images, the exact prompts, response parsing, and repetition, the result cannot be verified, and the improvement could be due to prompt sensitivity rather than ReFrame.
minor comments (5)
  1. [§5.1, Tables 1-3] The text in Section 5.1 refers to Table 2 when presenting completeness and Table 3 when presenting inconsistency, but the table contents appear swapped relative to those references; please check the table numbering and the column headers.
  2. [Table 1] The row labeled 'Ground Truth (validation set)' reports an inconsistency of 0.304 and completeness of 0.348, which is lower than the rectified model's completeness; this is surprising and should be explained, since ground-truth captions would be expected to align better with human-recognized objects.
  3. [§4.3] There are several typos, including 'BLUE' for 'BLEU', 'Masked R-CNN' for 'Mask R-CNN' in Section 5, and 'valuse' for 'values' in Section 4.2.2; a careful proofread is needed.
  4. [§3.1] The caption tree generation is described at a high level; please provide the recursive LSTM update equations or pseudo-code so that the branching procedure can be reproduced.
  5. [§5.4, Table 4] The 'Prominence-based Completeness' filter is not formally defined; specify how the object area thresholds (5%, 10%, 15%) are computed and whether the filter is applied to the rectifier's detections or to the explanation text.

Circularity Check

3 steps flagged · score 8.0 of 10

Headline gains are an artifact of scoring rectified outputs against the same Mask R-CNN detections used to create them; independent evaluation is needed before crediting improvement.

  1. self definitional [Section 4.3.2, Eq. (2); Section 3.1]
    "To calculate the completeness score, we tally the instances where the image explaining model accurately identifies and describes objects that are also detected and confirmed by the rectifier as illustrated in Figure 7. This count is then normalized by the total number of objects identified by the rectifier ... Completeness = #({DetectedObjects}∩{ObjectsWithRectification}) / #({ObjectsWithRectification}) (2)"

    The metric's numerator is the intersection with ObjectsWithRectification (MRCNN detections) and its denominator is that same set. The rectification step in Section 3.1 replaces candidate tokens with 'the token with the highest probability of the decoder that exists in the objects recognized by MRCNN', thereby increasing the intersection by construction. Any base model post-processed this way will show higher completeness on Eq. (2) even if no semantic improvement occurred; the 81.81% headline therefore measures agreement with the rectifier, not image understanding.

  2. self definitional [Section 4.3.1, Eq. (1); Sections 3.1-3.3]
    "the inconsistency score is calculated by identifying the instances where the image explaining model ... asserts the presence of objects and the rectifier does not detect in the image. This is computed by getting the set difference between them. The frequency of these instances is then normalized by the total number of objects identified by the rectifier ... Inconsistency = #({DetectedObjects}−{ObjectsWithRectification}) / #({ObjectsWithRectification}) (1)"

    Eq. (1) scores as 'inconsistent' exactly those detected objects outside the MRCNN detection set. The rectifier removes or replaces such out-of-set tokens, so the set difference shrinks by construction. The reported 37.10% inconsistency reduction is a tautological consequence of filtering with the same set used to define the metric; it does not require the corrected explanation to be more accurate.

1 more flagged steps
  1. fitted input called prediction [Section 4.2.3; Section 5.3]
    "the rectification prompts by the outputs of MRCNN, were manually fed into the GPT-4 model. By mentioning the '<inconsistent objects> does not exist in the image', and then asked the same question again."

    For the GPT-4 arm, the MRCNN detections are manually inserted into the prompt and then the output is scored by Eqs. (1)-(2), whose reference set is the same MRCNN detection set. The resulting completeness/inconsistency numbers are therefore forced agreement with the injected detections; the 5.2% inconsistency improvement and 0.02% completeness change are not independent evidence of better explanation.

full rationale

The core evaluation is circular by definition. Every headline gain is produced by making the model output agree with the MRCNN detection set, and every headline metric is defined relative to that exact set. No independent ground truth or human evaluation is used; Section 4.3 explicitly rejects conventional metrics because 'our proposed model's rectification goes beyond the provided ground truth.' The acknowledged limitation (baseline and rectifier trained on the same dataset) does not mitigate the metric circularity. The framework is self-contained against external benchmarks only in the sense that it invents its own benchmarks; the quantitative evidence as reported is forced. Score 8 rather than 10 because the underlying idea of using a detector to catch object-level errors is not logically impossible and the qualitative claim has independent content, but the numerical claims reduce to the definition of the metrics.

Assumptions & free parameters 1 free parameters · 4 assumptions · 0 invented entities

The paper relies on the assumptions that MRCNN detections are a true reference, that word2vec can bridge the vocabulary gap, and that the caption tree explores sufficiently. There are no fitted parameters reported, but the MRCNN detection threshold and top-k choices are unstated. No new entities are introduced.

free parameters (1)
  • MRCNN detection confidence threshold = not reported
    The confidence threshold used to accept a Mask R-CNN detection is not given in the paper. This threshold directly controls the set of 'Objects with Rectification' that appears in both the rectification step and the evaluation metrics.
assumptions (4)
  • domain assumption Mask R-CNN detections over 80 COCO classes are a correct and complete oracle for objects in an image.
    Used in Section 4.3 as the reference set for both metrics and in Section 3 to select replacement tokens.
  • domain assumption Word2vec embeddings can map decoder vocabulary tokens to the correct COCO class for any object word.
    Invoked in Section 3 to bridge the vocabulary gap between the caption decoder and the 80-class detector.
  • domain assumption The recursive LSTM caption tree explores enough candidate captions to include the objects the rectifier later keeps.
    The caption tree in Section 3.1 and Figure 5 is the only source of candidate tokens; if a true object is absent from all branches, the rectifier cannot add it.
  • domain assumption GPT-4 will incorporate re-prompted detector feedback into its revision.
    The prompt-based experiment in Section 4.2.3 manually feeds detector findings back and assumes GPT-4 revises its answer consistently.

how reviews work

0 comments
Cite this review

Pith. "Pith review of ReFrame: Rectification Framework for Image Explaining Architectures." pith.science (2026). https://pith.science/paper/IZ2FWKVG

@misc{pith2026250618272,
  author       = {Pith},
  title        = {Pith review of: ReFrame: Rectification Framework for Image Explaining Architectures},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/IZ2FWKVG}},
  note         = {Machine review of arXiv:2506.18272}
}
read the original abstract

Image explanation has been one of the key research interests in the Deep Learning field. Throughout the years, several approaches have been adopted to explain an input image fed by the user. From detecting an object in a given image to explaining it in human understandable sentence, to having a conversation describing the image, this problem has seen an immense change throughout the years, However, the existing works have been often found to (a) hallucinate objects that do not exist in the image and/or (b) lack identifying the complete set of objects present in the image. In this paper, we propose a novel approach to mitigate these drawbacks of inconsistency and incompleteness of the objects recognized during the image explanation. To enable this, we propose an interpretable framework that can be plugged atop diverse image explaining frameworks including Image Captioning, Visual Question Answering (VQA) and Prompt-based AI using LLMs, thereby enhancing their explanation capabilities by rectifying the incorrect or missing objects. We further measure the efficacy of the rectified explanations generated through our proposed approaches leveraging object based precision metrics, and showcase the improvements in the inconsistency and completeness of image explanations. Quantitatively, the proposed framework is able to improve the explanations over the baseline architectures of Image Captioning (improving the completeness by 81.81% and inconsistency by 37.10%), Visual Question Answering(average of 9.6% and 37.10% in completeness and inconsistency respectively) and Prompt-based AI model (0.01% and 5.2% for completeness and inconsistency respectively) surpassing the current state-of-the-art by a substantial margin.

Figures

Figures reproduced from arXiv: 2506.18272 by the authors.

Figure 2
Figure 2. Improved Explainability through our Proposed [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗
Figure 1
Figure 1. Shortcomings of Baseline Multi-modal Architec [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 3
Figure 3. Instances explaining the architecture and its improvement over baseline explanations. The table illustrates the [PITH_FULL_IMAGE:figures/full_fig_p003_3.png] view at source ↗
Figures from the paper (4 more)
Figure 4
Figure 4. Figure 4: Architecture of our proposed model, ReFrame [PITH_FULL_IMAGE:figures/full_fig_p005_4.png]
Figure 5
Figure 5. Figure 5: Caption Tree Generation over baseline Image Captioning [27] [PITH_FULL_IMAGE:figures/full_fig_p005_5.png]
Figure 6
Figure 6. Figure 6: Rectification over baseline VQA model. leveraging the strengths of GPT-4’s generative capabilities and MRCNN’s precise object detection to produce more consistent and complete image explanations. 4 Implementation Details and Evaluation Metrics 4.1 Dataset Description I…
Figure 7
Figure 7. Figure 7: Inconsistency and Completeness over sample Input Image. The image in left represents the classes identified by any [PITH_FULL_IMAGE:figures/full_fig_p007_7.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

33 extracted references · 16 canonical work pages

  1. [7]

    Guowei Ge, Yufeng Han, Lingguang Hao, Kuangrong Hao, Bing Wei, and Xue- song Tang. 2024. Show, tell and rectify: Boost image caption generation via an output rectifier. Neurocomputing (2024), 127651

    work page 2024

  2. [1]

    Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Floren- cia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. 2023. Gpt-4 technical report. arXiv preprint arXiv:2303.08774 (2023)

    arXiv 2023

  3. [2]

    Stanislaw Antol, Aishwarya Agrawal, Jiasen Lu, Margaret Mitchell, Dhruv Batra, C Lawrence Zitnick, and Devi Parikh. 2015. Vqa: Visual question answering. In Proceedings of the IEEE international conference on computer vision . 2425–2433

    work page 2015

  4. [3]

    Zechen Bai, Pichao Wang, Tianjun Xiao, Tong He, Zongbo Han, Zheng Zhang, and Mike Zheng Shou. 2024. Hallucination of multimodal large language models: A survey. arXiv preprint arXiv:2404.18930 (2024)

    arXiv 2024

  5. [4]

    Zhixuan Chu, Mengxuan Hu, Qing Cui, Longfei Li, and Sheng Li. 2024. Task- driven causal feature distillation: Towards trustworthy risk prediction. In Pro- ceedings of the AAAI Conference on Artificial Intelligence , Vol. 38. 11642–11650

    work page 2024

  6. [5]

    So what if ChatGPT wrote it?

    Yogesh K Dwivedi, Nir Kshetri, Laurie Hughes, Emma Louise Slade, Anand Je- yaraj, Arpan Kumar Kar, Abdullah M Baabdullah, Alex Koohang, Vishnupriya Raghavan, Manju Ahuja, et al. 2023. “So what if ChatGPT wrote it?” Multidisci- plinary perspectives on opportunities, challenges and implications of generative conversational AI for research, practice and poli...

    work page 2023

  7. [6]

    Ali Farhadi, Mohsen Hejrati, Mohammad Amin Sadeghi, Peter Young, Cyrus Rashtchian, Julia Hockenmaier, and David Forsyth. 2010. Every picture tells a story: Generating sentences from images. In Computer Vision–ECCV 2010: 11th European Conference on Computer Vision, Heraklion, Crete, Greece, September 5-11, 2010, Proceedings, Part IV 11 . Springer, 15–29

    work page 2010

  8. [8]

    Yunchao Gong, Liwei Wang, Micah Hodosh, Julia Hockenmaier, and Svetlana Lazebnik. 2014. Improving image-sentence embeddings using large weakly annotated photo collections. In Computer Vision–ECCV 2014: 13th European Con- ference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part IV 13 . Springer, ReFrame: Rectification Framework for Image Expla...

    work page 2014

Show all 33 references
  1. [9]

    Kaiming He, Georgia Gkioxari, Piotr Dollár, and Ross Girshick. 2017. Mask r-cnn. In Proceedings of the IEEE international conference on computer vision . 2961–2969

    2017

  2. [10]

    Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. 2016. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition . 770–778

    2016

  3. [11]

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

  4. [12]

    Yushi Hu, Hang Hua, Zhengyuan Yang, Weijia Shi, Noah A Smith, and Jiebo Luo. 2023. Promptcap: Prompt-guided image captioning for vqa with gpt-3. In Proceedings of the IEEE/CVF International Conference on Computer Vision . 2963– 2975

    2023

  5. [13]

    Wen Huang, Hongbin Liu, Minxin Guo, and Neil Zhenqiang Gong. 2024. Vi- sual hallucinations of multi-modal large language models. arXiv preprint arXiv:2402.14683 (2024)

    2024 arXiv

  6. [14]

    Wonjae Kim, Bokyung Son, and Ildoo Kim. 2021. Vilt: Vision-and-language trans- former without convolution or region supervision. In International conference on machine learning. PMLR, 5583–5594

    2021

  7. [15]

    Ryan Kiros, Ruslan Salakhutdinov, and Rich Zemel. 2014. Multimodal neural language models. InInternational conference on machine learning. PMLR, 595–603

    2014

  8. [16]

    Chenliang Li, Haiyang Xu, Junfeng Tian, Wei Wang, Ming Yan, Bin Bi, Jiabo Ye, Hehong Chen, Guohai Xu, Zheng Cao, et al. 2022. mplug: Effective and effi- cient vision-language learning by cross-modal skip-connections. arXiv preprint arXiv:2205.12005 (2022)

    2022 arXiv

  9. [17]

    Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dollár, and C Lawrence Zitnick. 2014. Microsoft coco: Common objects in context. In Computer Vision–ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proce...

    2014

  10. [18]

    Annika Lindh, Robert Ross, and John D Kelleher. 2023. Show, Prefer and Tell: Incorporating User Preferences into Image Captioning. In Proceedings of the 38th ACM/SIGAPP Symposium on Applied Computing . 1139–1142

    2023

  11. [19]

    Tomas Mikolov, Kai Chen, Greg Corrado, and Jeffrey Dean. 2013. Efficient estimation of word representations in vector space.arXiv preprint arXiv:1301.3781 (2013)

    2013 arXiv

  12. [20]

    Tomas Mikolov, Ilya Sutskever, Kai Chen, Greg S Corrado, and Jeff Dean. 2013. Distributed representations of words and phrases and their compositionality. Advances in neural information processing systems 26 (2013)

    2013

  13. [21]

    Jiquan Ngiam, Aditya Khosla, Mingyu Kim, Juhan Nam, Honglak Lee, and An- drew Y Ng. 2011. Multimodal deep learning. InProceedings of the 28th international conference on machine learning (ICML-11) . 689–696

    2011

  14. [22]

    Khanh Nguyen, Ali Furkan Biten, Andres Mafla, Lluis Gomez, and Dimosthenis Karatzas. 2023. Show, interpret and tell: entity-aware contextualised image captioning in wikipedia. In Proceedings of the AAAI Conference on Artificial Intelligence, Vol. 37. 1940–1948

    2023

  15. [23]

    Xuran Pan, Tianzhu Ye, Dongchen Han, Shiji Song, and Gao Huang. 2022. Con- trastive language-image pre-training with knowledge graphs. Advances in Neural Information Processing Systems 35 (2022), 22895–22910

    2022

  16. [24]

    Thorsten Rudroff. 2024. Revealing the Complexity of Fatigue: A Review of the Persistent Challenges and Promises of Artificial Intelligence. Brain sciences 14, 2 (2024), 186

    2024

  17. [25]

    Yongliang Shen, Kaitao Song, Xu Tan, Dongsheng Li, Weiming Lu, and Yueting Zhuang. 2024. Hugginggpt: Solving ai tasks with chatgpt and its friends in hugging face. Advances in Neural Information Processing Systems 36 (2024)

    2024

  18. [26]

    Matteo Stefanini, Marcella Cornia, Lorenzo Baraldi, Silvia Cascianelli, Giuseppe Fiameni, and Rita Cucchiara. 2022. From show to tell: A survey on deep learning- based image captioning. IEEE transactions on pattern analysis and machine intelligence 45, 1 (2022), 539–559

    2022

  19. [27]

    Oriol Vinyals, Alexander Toshev, Samy Bengio, and Dumitru Erhan. 2015. Show and tell: A neural image caption generator. In Proceedings of the IEEE conference on computer vision and pattern recognition . 3156–3164

    2015

  20. [28]

    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

  21. [29]

    Yiqi Wang, Wentao Chen, Xiaotian Han, Xudong Lin, Haiteng Zhao, Yongfei Liu, Bohan Zhai, Jianbo Yuan, Quanzeng You, and Hongxia Yang. 2024. Exploring the reasoning abilities of multimodal large language models (mllms): A com- prehensive survey on emerging trends in multimodal ...

    2024 arXiv

  22. [30]

    Zichao Yang, Xiaodong He, Jianfeng Gao, Li Deng, and Alex Smola. 2016. Stacked attention networks for image question answering. In Proceedings of the IEEE conference on computer vision and pattern recognition . 21–29

    2016

  23. [31]

    Zhou Yu, Jun Yu, Yuhao Cui, Dacheng Tao, and Qi Tian. 2019. Deep modular co- attention networks for visual question answering. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition . 6281–6290

    2019

  24. [32]

    Rui Zhao, Ruqiang Yan, Zhenghua Chen, Kezhi Mao, Peng Wang, and Robert X Gao. 2019. Deep learning and its applications to machine health monitoring. Mechanical Systems and Signal Processing 115 (2019), 213–237

    2019

  25. [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

Pith tools

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