Pith. sign in

REVIEW 4 major objections 6 minor 44 references

VCRScore: Image captioning metric based on V\&L Transformers, CLIP, and precision-recall

T0 review · 4 major / 6 minor · reviewed 2026-08-10 · deepseek-v4-flash

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

Pith's one-line read This paper claims that VCRScore, a supervised metric combining CLIP retrieval embeddings, a vision-language transformer distance, and precision-recall features, agrees with human quality ratings better than existing image-captioning…

desk verdict A useful but flawed attempt at a learned captioning metric; the evaluation is compromised by image-level split leakage and zero-score removal, so the central correlation claim is not yet established. read the letter →

arxiv 2501.09155 v2 pith:6LIDZBER submitted 2025-01-15 cs.CV cs.CL

Guillermo Ruiz , Tania Ramírez , Daniela Moctezuma This is my paper · ORCID
classification cs.CVcs.CL
keywords imagecaptioningevaluationVCRScorevision-languagetransformerCLIPprecision-recallhumancorrelationgradientboostingSpearman
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

This paper tries to establish that image-captioning evaluation can be improved by replacing n-gram and raw embedding metrics with a learned regressor trained on human quality scores. The authors built a 600-pair human-tagged dataset from MS COCO images and six caption sources, then trained VCRScore on those scores plus four public datasets, using features from CLIP-based retrieval models, a vision-language transformer, and lexical precision and recall. Reported Spearman correlations with human ratings are 0.649 on 360 held-out pairs and 0.743 on 17,794 external test pairs, higher than every compared metric. If this holds, captioning researchers get a metric that ranks models closer to human judgment than BLEU, CIDEr, BertScore, or CLIPScore.

What carries the argument

The load-bearing object is the VCRScore regressor: a Gradient Boosting Regressor that maps image-caption features into a 0-to-1 human-like score. Its features are cosine similarities from CLIP and retrieval-tuned CLIP between the image, the caption, and references; a distance from the ViLT vision-language transformer; and precision and recall computed as the proportion of caption words found in a pool formed by reference words plus object names detected by YOLOv3 and DETR. Precision and recall supply lexical coverage, CLIP supplies semantic alignment, and ViLT supplies multimodal alignment; the regressor learns how to combine them to match human ratings.

What would settle it

Re-run the same VCRScore training and the Table 11 comparison on the full external dataset without deleting the 21,974 zero-score samples, comparing Spearman correlations on all 81,289 pairs; if VCRScore no longer beats CLIPScore ref and MCIPScore ref, the paper's central claim fails. A second check is to train on some datasets and test on held-out datasets from unseen sources to see whether the advantage survives cross-domain transfer.

Watch

Extended reading notes

Core claim

In the paper's own terms, the discovery is that a trainable metric can be highly correlated with human tagging: VCRScore reaches the highest Spearman correlations among all metrics tested, 0.649 on the authors' held-out 360 image-caption pairs and 0.743 on 17,794 external test pairs. The metric is trained on the mean of eight human ratings per pair, using a Gradient Boosting Regressor whose features are CLIP and retrieval-tuned CLIP similarities, a ViLT image-caption distance, and precision and recall computed against a pool of reference words augmented by object detectors. The authors interpret this as evidence that semantic embedding signals and lexical coverage are complementary, and that a supervised combination of them approximates human judgment better than any single existing metric.

Load-bearing premise

The comparison stands on treating the removal of all 21,974 zero-scored external samples as harmless; if low-quality pairs were kept, the reported correlation advantage could shrink or reverse.

Editorial extensions

If this is right

  • If the reported correlations hold, model rankings from VCRScore are a closer surrogate for human preference than rankings from BLEU, ROUGE, METEOR, SPICE, CIDEr, BertScore, CLIPScore, or MCIPScore.
  • The result supports combining lexical coverage with semantic and multimodal alignment rather than relying on any single signal.
  • VCRScore's 0-to-1 output on the human scale makes it usable as a quality filter for caption outputs, unlike unbounded lexical metrics.
  • The reported advantage is limited to test splits drawn from the same datasets used for training after zero-score samples were removed.

Reading between the lines

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

  • If zero-scored pairs were kept in training and testing, the reported gap would likely narrow because low-quality pairs are where lexical and semantic metrics fail; this is a testable prediction, not a paper claim.
  • Because tagger agreement is only moderate, part of VCRScore's higher correlation may reflect fitting annotator noise; averaging more taggers per pair would test that.
  • The same feature recipe could be evaluated reference-free by replacing reference words in the pool with detected objects alone, a variant the paper does not report.
  • A direct extension would swap ViLT for a video-language transformer and object detectors for temporal detections to score video captions; no evidence here supports that transfer.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 6 minor

Summary. The paper proposes VCRScore, a learned image-captioning metric that combines ViLT similarity, CLIP/MCIP variants (MCIPScore, MCIPScore ref), and precision/recall features computed against a pool of reference words and object-detection labels, into a Gradient Boosting Regressor trained to predict human ratings. The authors construct a new human-labeled dataset of 600 image-caption pairs from 100 COCO images, score them with four taggers on a five-point scale, and also use external datasets (VICR, Flickr8k-Expert, Flickr8k-CF, Composite). After normalizing scores and excluding all zero-score external samples, they train on 70% of external data and 30% of their own data, and test on the remaining splits. They report Spearman correlations with human scores (Table 11: 0.649 on their own test set and 0.743 on the external test set) and conclude that VCRScore is the most human-correlated metric among those compared.

Significance. If the reported correlations were robust, VCRScore would be a useful addition to the image-captioning evaluation toolbox, and the new human-labeled dataset could be a resource for the community. The paper is transparent about its data-processing choices, compares a wide range of lexical and semantic metrics, and provides code via GitHub. However, three load-bearing limitations prevent the paper from establishing its central claim: (1) the exclusion of all zero-score samples truncates the evaluation distribution; (2) the train/test splits are not shown to be image-disjoint, creating a risk of image-level leakage; and (3) the main comparison pits a supervised regressor trained in-domain against zero-shot baselines, so the reported advantage is partly a measure of fit to the annotation distribution. These issues are fixable with additional experiments, but the current evidence is insufficient.

major comments (4)
  1. [Section 4.1, Table 8] The paper removes all 21,974 samples with human score 0 from the external datasets before splitting into train/test, reducing the data from 81,289 to 59,315. The stated reason ('not helpful to our model generation') is not a valid justification for excluding the low-quality end of the rating scale from the test set. Because the test set in Table 11 is therefore composed entirely of captions with positive human scores, the reported Spearman correlations measure agreement only over the upper part of the quality range. This is precisely the range where lexical and embedding metrics are most competitive; the metrics' known failure mode is on low-quality captions. To support the central claim, the authors must report correlations with zero-score samples included (at least in the test set), or justify why those samples should be excluded from evaluation for all metrics.
  2. [Section 4.1, Table 8] The split is not described as stratified by image. The authors' own dataset consists of 600 image-caption pairs from 100 images (six captions per image: five model outputs plus the COCO reference). A 30/70 pair-level split will almost certainly place different captions of the same image in both train and test. Since VCRScore's feature stack includes ViLT and CLIP/MCIP image-text similarities, a gradient-boosting regressor can memorize image-specific visual patterns and transfer them to held-out captions of the same image, inflating the 0.649 Spearman correlation on the own test set. The external dataset split raises the same concern for datasets with multiple captions per image. The authors should run an image-disjoint split (no image appears in both train and test) and report the resulting correlations; if they have already done this, the paper needs to state it.
  3. [Section 4.3, Table 11] The headline comparison is between a supervised regressor (VCRScore) trained on the human-score distribution of the test datasets and zero-shot baselines (BLEU, ROUGE, METEOR, CIDEr, SPICE, BertScore, CLIPScore, MCIPScore, ViLT) that have not been trained on those labels. The regressor's objective is to predict the very scores against which it is later evaluated; the held-out split mitigates but does not remove the in-domain training advantage. The reported correlations therefore do not establish that VCRScore is a better metric; they establish that a regressor fit to a dataset can rank that dataset's labels. To make the comparison meaningful, the authors should include (a) cross-dataset evaluation (train on one or more external datasets, test on others), and (b) a trained baseline that uses the same regression algorithm on the baseline metric features (e.g., GBR trained to predict human scores from CLIPScore, BertScore, etc.). Without these controls, the claimed superiority in Table 11 is not interpretable.
  4. [Section 3.2, Eqs. (4)-(6)] The precision and recall features depend on the construction of the word 'pool' from the references and object detection outputs. The paper does not specify which YOLOv3/DETR versions and confidence thresholds were used, how object labels are mapped to words (e.g., singular/plural, synonyms), or how stopwords are handled in the caption tokens. It also does not report an ablation that quantifies the contribution of precision and recall to the final VCRScore correlation, despite these features being a stated part of the proposal. Without this information, the proposed metric is not fully reproducible and the role of the precision-recall component is unverified.
minor comments (6)
  1. [Section 2.1] The sentence 'we have a set of 600 scores produced twice by the same tagger' is confusing because there are four taggers; it should say 'each tagger produced 600 scores in each of two stages.'
  2. [Section 3.2] The text says 'Both values are between 0 and 1' for precision and recall, but Equations (5) and (6) multiply by 100, so the values are percentages between 0 and 100.
  3. [Table 11] The table reports only point estimates of Spearman correlations. With 360 and 17,794 test samples, the differences between adjacent entries (e.g., 0.743 vs. 0.716) may not be statistically significant; the authors should provide confidence intervals or a significance test.
  4. [Section 4.1] The asymmetric split (30/70 for the own dataset, 70/30 for the external data) is justified only by 'because of its size.' This is not an explanation; the authors should state the rationale or use a consistent split strategy.
  5. [Figure 9] The caption 'Correlations on positions' is vague; it should specify that the values are Spearman correlations between the model rankings produced by each metric and the human-generated ranking.
  6. [References] References [22] and [23] are duplicates of the same BLIP-2 paper, and references [33] and [34] both cite the CLIP paper; one of each pair should be removed or consolidated.

Circularity Check

1 steps flagged · score 4.0 of 10

No definitional circularity in the supervised-target design, but the unreported image-level split creates a same-image leakage path that partly forces VCRScore's headline correlations.

  1. fitted input called prediction [Section 4.1 / Table 8 (split); Section 3.1.2 (ViLT/MCIP features); Section 4.3 / Table 11 (reported VCRScore correlation)]
    "Our build dataset was split into 30-70 percent of train and test, respectively. In the case of the other datasets, as a whole, it was split into 70-30 percent train and test, respectively. ... For this work, the ViLT obtains a distance score between the model encoded by the image and the caption."

    VCRScore's regressor uses image-dependent features (ViLT image-caption distance and CLIP/MCIP similarities), and the datasets contain multiple captions per image (own set: 100 images x 6 captions = 600 pairs; external sets have many captions per image). The split is described only as row percentages with no image-grouping or image-overlap check. Under that row-level split, captions of the same image occur in both training and test partitions, so a tree ensemble can learn image-specific score patterns from training captions and apply them to test captions of the same image.

full rationale

There is no definitional circularity: the GBR is trained on a split and evaluated on held-out rows, so a supervised metric whose target is human score is not itself circular. The self-citations ([12], [28]) are not load-bearing: [12] only motivates BertGrammar, a separate baseline, and [28] is a review reference; neither forces VCRScore's result. The precision/recall and ViLT/CLIP features are defined independently of the target. The main circularity-adjacent problem is the split construction: the paper reports only 30/70 and 70/30 row splits and never states that the split is stratified by image, despite multi-caption-per-image datasets and image-dependent features. If the split is row-level, the headline correlations are partly attributable to same-image fitting, which is the specific reduction quoted above. The zero-score removal (Section 4.1) is a distributional bias that should be reported as a validity threat, but it is not a circular reduction. Given that the rest of the derivation is standard held-out supervised evaluation, a moderate score of 4 is appropriate; a higher score would require the image overlap to be confirmed in the released code.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

The central claim rests on a trained regression model and several data-processing choices that are not fully specified: the GBR hyperparameters, normalization of external scores, and object detection thresholds are all absent from the manuscript. The metric itself is a statistical model of human ratings, so the reported correlations are partly a measure of the fit rather than an independent discovery. There are no invented physical or conceptual entities; the assumptions are about the adequacy of the human labels, the pool of words used for recall, the interval nature of the rating scale, and the representativeness of the 100-image sample.

free parameters (3)
  • Gradient Boosting Regressor hyperparameters (n_estimators, learning_rate, max_depth, feature subset) = Not reported
    The GBR is the trained model at the core of VCRScore; without its hyperparameters and the final chosen feature list the reported correlations cannot be reproduced or independently checked. These choices were presumably made during development rather than derived from theory.
  • Score normalization scheme for external datasets = Not reported
    The paper states the scores were normalized to [0,1] but does not give the mapping (e.g., min-max versus divide-by-max) for each of VICR, Flickr8k-Expert, Flickr8k-CF, Composite. The resulting labels are the regression target and affect all correlations.
  • Object detection model versions and confidence thresholds for pool construction = Not reported
    The 'pool' is built from YOLOv3 and DETR object labels plus references; the specific model snapshots and any filtering thresholds change recall and precision features, hence the metric's output.
assumptions (4)
  • domain assumption Human scores on the five-point 0,0.25,...,1 scale are an interval scale, so arithmetic means of tagger scores are valid targets for regression.
    Section 2.2: the paper averages the eight tagger evaluations per pair and uses the mean as the regression target y. Treating ordinal Likert-style anchors as equal-interval numbers is an assumption, especially with one tagger of low reliability.
  • ad hoc to paper The 'pool' of reference words plus YOLOv3/DETR object labels captures the salient content of the image, so recall against this pool is a meaningful measure of information coverage.
    Section 3.2, Eqs. (4)-(6): recall is defined as |Z∩pool|/|pool|; if the object detectors miss salient objects or add spurious labels, the recall feature does not measure what the metric claims.
  • ad hoc to paper Excluding all samples with human score 0 from the external datasets is harmless and improves model generation.
    Section 4.1: 21,974 zero-score samples are removed. This truncates the label distribution for both training and testing and is a post hoc choice with no external justification.
  • domain assumption The 100 MS-COCO images and the six caption sources (five models plus one human reference) are representative enough to train and validate a general-purpose captioning metric.
    Section 2: the entire in-house dataset is 600 pairs from 100 images; the paper uses this small set as a central part of the evaluation, so the assumption that it represents the broader captioning distribution is load-bearing.

how reviews work

0 comments
Cite this review

Pith. "Pith review of VCRScore: Image captioning metric based on V\&L Transformers, CLIP, and precision-recall." pith.science (2026). https://pith.science/paper/6LIDZBER

@misc{pith2026250109155,
  author       = {Pith},
  title        = {Pith review of: VCRScore: Image captioning metric based on V\&L Transformers, CLIP, and precision-recall},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/6LIDZBER}},
  note         = {Machine review of arXiv:2501.09155}
}
read the original abstract

Image captioning has become an essential Vision & Language research task. It is about predicting the most accurate caption given a specific image or video. The research community has achieved impressive results by continuously proposing new models and approaches to improve the overall model's performance. Nevertheless, despite increasing proposals, the performance metrics used to measure their advances have remained practically untouched through the years. A probe of that, nowadays metrics like BLEU, METEOR, CIDEr, and ROUGE are still very used, aside from more sophisticated metrics such as BertScore and ClipScore. Hence, it is essential to adjust how are measure the advances, limitations, and scopes of the new image captioning proposals, as well as to adapt new metrics to these new advanced image captioning approaches. This work proposes a new evaluation metric for the image captioning problem. To do that, first, it was generated a human-labeled dataset to assess to which degree the captions correlate with the image's content. Taking these human scores as ground truth, we propose a new metric, and compare it with several well-known metrics, from classical to newer ones. Outperformed results were also found, and interesting insights were presented and discussed.

Figures

Figures reproduced from arXiv: 2501.09155 by the authors.

Figure 1
Figure 1. Screenshot of the simple tagging designed tool [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. Overall score for each model, as well as the Reference of the MS [PITH_FULL_IMAGE:figures/full_fig_p009_2.png] view at source ↗
Figure 3
Figure 3. Scores assigned to each model’s prediction by all the taggers in the [PITH_FULL_IMAGE:figures/full_fig_p009_3.png] view at source ↗
Figures from the paper (7 more)
Figure 4
Figure 4. Figure 4: Examples of captions in the different datasets. (a) is for VICR, (b) is [PITH_FULL_IMAGE:figures/full_fig_p016_4.png]
Figure 5
Figure 5. Figure 5: HeatMap of the correlation results of all the metrics evaluated [PITH_FULL_IMAGE:figures/full_fig_p019_5.png]
Figure 6
Figure 6. Figure 6: Scores of the models with proposed metric and the human score [PITH_FULL_IMAGE:figures/full_fig_p020_6.png]
Figure 7
Figure 7. Figure 7: Scores of the models with the human score (Tagging) compared to [PITH_FULL_IMAGE:figures/full_fig_p021_7.png]
Figure 8
Figure 8. Figure 8: Scores of the models with the human score (Tagging) compared to [PITH_FULL_IMAGE:figures/full_fig_p022_8.png]
Figure 9
Figure 9. Figure 9: Correlations on positions between the metrics and the tagging process [PITH_FULL_IMAGE:figures/full_fig_p023_9.png]
Figure 10
Figure 10. Figure 10: How each metric ranks all the models usage of embeddings or models not currently used as performance metrics had a very outstanding performance; for instance, the adaptation of the version of CLIP for retrieval option into the ClipScore metric significantly improved t…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

44 extracted references · 39 canonical work pages

  1. [1]

    Springer New York, New York, NY, 2008

    Spearman Rank Correlation Coefficient, pages 502–505. Springer New York, New York, NY, 2008

    work page 2008

  2. [2]

    A. N. Abu-Safe and S. E. Elrofai. Qos–aware meta-heuristic services se- lection algorithm and likert scale measurement for iot environment. In- ternational Journal of Computer Science Trends and Technology (IJCST) , 8(1):1–8, 2020

    work page 2020

  3. [3]

    Aditya, Y

    S. Aditya, Y. Yang, C. Baral, Y. Aloimonos, and C. Ferm¨ uller. Image understanding using vision and reasoning through scene description graph. Computer Vision and Image Understanding , 173:33–45, 2018

    work page 2018

  4. [4]

    Anderson, B

    P. Anderson, B. Fernando, M. Johnson, and S. Gould. Spice: Semantic propositional image caption evaluation. In European conference on com- puter vision , pages 382–398, Amsterdam, Netherlands, 2016. Springer

    work page 2016

  5. [5]

    Aneja, A

    J. Aneja, A. Deshpande, and A. Schwing. Convolutional image captioning. In Computer Vision and Pattern Recognition, pages 5561–5570, Utah, USA,

    work page

  6. [6]

    Carion, F

    N. Carion, F. Massa, G. Synnaeve, N. Usunier, A. Kirillov, and S. Zagoruyko. End-to-end object detection with transformers. In Euro- pean conference on computer vision , pages 213–229. Springer, 2020

    work page 2020

  7. [7]

    Cornia, M

    M. Cornia, M. Stefanini, L. Baraldi, and R. Cucchiara. Meshed-Memory Transformer for Image Captioning. In Proceedings of the IEEE/CVF Con- ference on Computer Vision and Pattern Recognition , pages 10578–10587, Virtual, 2020. IEEE/CVF

    work page 2020

  8. [8]

    Y. Cui, G. Yang, A. Veit, X. Huang, and S. Belongie. Learning to evaluate image captioning. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) , June 2018

    work page 2018

Show all 44 references
  1. [9]

    Devlin, M

    J. Devlin, M. W. Chang, K. Lee, and K. Toutanova. BERT: Pre-training of deep bidirectional transformers for language understanding. NAACL HLT 2019 - 2019 Conference of the North American Chapter of the As- sociation for Computational Linguistics: Human Language Technologies - ...

    2019

  2. [10]

    S. M. Downing. Reliability: on the reproducibility of assessment data. Medical education, 38(9):1006–1012, 2004

    2004

  3. [11]

    J. H. Friedman. Greedy function approximation: a gradient boosting ma- chine. Annals of statistics , pages 1189–1232, 2001

    2001

  4. [12]

    Gonz´ alez-Ch´ avez, G

    O. Gonz´ alez-Ch´ avez, G. Ruiz, D. Moctezuma, and T. Ramirez-delReal. Are metrics measuring what they should? an evaluation of image captioning task metrics. Signal Processing: Image Communication , 120:117071, 2024

    2024

  5. [13]

    X. Han, J. Chang, and K. Wang. You only look once: unified, real-time object detection. Procedia Computer Science, 183(1):61–72, 2021

    2021

  6. [14]

    Hodosh, P

    M. Hodosh, P. Young, and J. Hockenmaier. Framing image description as a ranking task: data, models and evaluation metrics. J. Artif. Int. Res. , 47(1):853–899, may 2013

    2013

  7. [15]

    Kasai, K

    J. Kasai, K. Sakaguchi, L. Dunagan, J. Morrison, R. Le Bras, Y. Choi, and N. A. Smith. Transparent human evaluation for image captioning. In M. Carpuat, M.-C. de Marneffe, and I. V. Meza Ruiz, editors,Proceedings of the 2022 Conference of the North American Chapter of the Asso...

    2022

  8. [16]

    M. G. Kendall. Rank correlation methods. 1948

    1948

  9. [17]

    W. Kim, B. Son, and I. Kim. Vilt: Vision-and-language transformer with- out convolution or region supervision. In International conference on ma- chine learning, pages 5583–5594. PMLR, 2021. 25

    2021

  10. [18]

    Krippendorff

    K. Krippendorff. Testing the reliability of content analysis data. The content analysis reader, pages 350–357, 2009

    2009

  11. [19]

    Krippendorff

    K. Krippendorff. Agreement and information in the reliability of coding. Communication methods and measures , 5(2):93–112, 2011

    2011

  12. [20]

    Lavie and A

    A. Lavie and A. Agarwal. METEOR: An automatic metric for MT evalua- tion with high levels of correlation with human judgments. Proceedings of the Second Workshop on Statistical Machine Translation , (June):228–231, 2007

    2007

  13. [21]

    Levinboim, A

    T. Levinboim, A. V. Thapliyal, P. Sharma, and R. Soricut. Quality es- timation for image captions based on large-scale human evaluations. In K. Toutanova, A. Rumshisky, L. Zettlemoyer, D. Hakkani-Tur, I. Beltagy, S. Bethard, R. Cotterell, T. Chakraborty, and Y. Zhou, editors, ...

    2021

  14. [22]

    J. Li, D. Li, S. Savarese, and S. Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models, 2023

    2023

  15. [23]

    J. Li, D. Li, S. Savarese, and S. Hoi. Blip-2: bootstrapping language- image pre-training with frozen image encoders and large language models. In Proceedings of the 40th International Conference on Machine Learning , ICML’23. JMLR.org, 2023

    2023

  16. [24]

    R. Likert. A technique for the measurement of attitudes. Archives of psychology, 1932

    1932

  17. [25]

    Lin and F

    C.-Y. Lin and F. J. Och. ORANGE: a method for evaluating automatic evaluation metrics for machine translation. In COLING 2004: Proceedings of the 20th International Conference on Computational Linguistics , pages 501–507, Geneva, Switzerland, aug 23–aug 27 2004. COLING

    2004

  18. [26]

    T.-Y. Lin, M. Maire, S. Belongie, J. Hays, P. Perona, D. Ramanan, P. Doll´ ar, and C. L. Zitnick. Microsoft coco: Common objects in con- text. In European conference on computer vision , pages 740–755, Zurich, Switzerland, 2014. Springer

    2014

  19. [27]

    Q. Ma, Y. Graham, S. Wang, and Q. Liu. Blend: a novel combined mt metric based on direct assessment—casict-dcu submission to wmt17 metrics task. In Proceedings of the second conference on machine translation, pages 598–603, 2017

    2017

  20. [28]

    Moctezuma, T

    D. Moctezuma, T. Ram ´ ırez-delReal, G. Ruiz, and O. Gonz´ alez-Ch´ avez. Video captioning: a comparative review of where we are and which could be the route. Computer Vision and Image Understanding , page 103671, 2023. 26

    2023

  21. [29]

    L. D. Narins, A. Scott, A. Gautam, A. Kulkarni, M. Castanon, B. Kao, S. Ihorn, Y.-T. Siu, J. M. Mason, A. Blum, and I. Yoon. Validated image caption rating dataset. In A. Oh, T. Naumann, A. Globerson, K. Saenko, M. Hardt, and S. Levine, editors, Advances in Neural Information ...

    2023

  22. [30]

    M. E. Otter, G. J. Mellenbergh, and K. d. Glopper. The relation between information-processing variables and test-retest stability for questionnaire items. Journal of Educational Measurement , 32(2):199–216, 1995

    1995

  23. [31]

    Papineni, S

    K. Papineni, S. Roukos, T. Ward, and W.-J. Zhu. Bleu: a method for auto- matic evaluation of machine translation. In Proceedings of the 40th annual meeting of the Association for Computational Linguistics , pages 311–318, Philadelphia, USA, 2002. Association for Computational ...

    2002

  24. [32]

    L. Puka. Kendall’s Tau, pages 713–715. Springer Berlin Heidelberg, Berlin, Heidelberg, 2011

    2011

  25. [33]

    Radford, J

    A. Radford, J. W. Kim, C. Hallacy, A. Ramesh, G. Goh, S. Agarwal, G. Sastry, A. Askell, P. Mishkin, J. Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748–8763. PMLR, 2021

    2021

  26. [34]

    Radford, J

    A. Radford, J. W. Kim, C. Hallacy, A. Ramesh, G. Goh, S. Agarwal, G. Sas- try, A. Askell, P. Mishkin, J. Clark, G. Krueger, and I. Sutskever. Learn- ing transferable visual models from natural language supervision. CoRR, abs/2103.00020:1–16, 2021

    2021 arXiv

  27. [35]

    Rahman, N

    M. Rahman, N. Mohammed, N. Mansoor, and S. Momen. Chittron: An automatic bangla image captioning system. Procedia Computer Science , 154:636–642, 2019

    2019

  28. [36]

    Schall, K

    K. Schall, K. U. Barthel, N. Hezel, and K. Jung. Optimizing CLIP mod- els for image retrieval with maintained joint-embedding alignment. In E. Ch´ avez, B. Kimia, J. Lokoc, M. Patella, and J. Sedmidubsk, editors, Similarity Search and Applications - 17th International Conferen...

    2024

  29. [37]

    Schuster, R

    S. Schuster, R. Krishna, A. Chang, L. Fei-Fei, and C. D. Manning. Generat- ing semantically precise scene graphs from textual descriptions for improved image retrieval. In Proceedings of the fourth workshop on vision and lan- guage, pages 70–80, Lisbon, Portugal, 2015. Associa...

    2015

  30. [38]

    Stanojevi´ c and K

    M. Stanojevi´ c and K. Sima’an. Beer: Better evaluation as ranking. In Pro- ceedings of the Ninth Workshop on Statistical Machine Translation , pages 414–419, 2014. 27

    2014

  31. [39]

    Vedantam, C

    R. Vedantam, C. Lawrence Zitnick, and D. Parikh. Cider: Consensus- based image description evaluation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4566–4575, Boston, Mas- sachusetts, USA, 2015. IEEE/CVF

    2015

  32. [40]

    P. Wang, A. Yang, R. Men, J. Lin, S. Bai, Z. Li, J. Ma, C. Zhou, J. Zhou, and H. Yang. Ofa: Unifying architectures, tasks, and modalities through a simple sequence-to-sequence learning framework. CoRR, abs/2202.03052, 2022

    2022 arXiv

  33. [41]

    W. Wang, Z. Chen, and H. Hu. Hierarchical attention network for image captioning. In Proceedings of the AAAI conference on artificial intelligence, volume 33, pages 8957–8964, 2019

    2019

  34. [42]

    Wiebe, R

    J. Wiebe, R. Bruce, and T. P. O’Hara. Development and use of a gold- standard data set for subjectivity classifications. In Proceedings of the 37th annual meeting of the Association for Computational Linguistics , pages 246–253, 1999

    1999

  35. [43]

    K. Xu, J. Ba, R. Kiros, K. Cho, A. Courville, R. Salakhudinov, R. Zemel, and Y. Bengio. Show, attend and tell: Neural image caption generation with visual attention. In F. Bach and D. Blei, editors, Proceedings of the 32nd International Conference on Machine Learning, volume 3...

    2015

  36. [44]

    Zhang*, V

    T. Zhang*, V. Kishore*, F. Wu*, K. Q. Weinberger, and Y. Artzi. Bertscore: Evaluating text generation with bert. In International Con- ference on Learning Representations , pages 1–41, Addis Ababa, Ethiopia,

Pith tools

Reviewed August 10, 2026 · model on record in the stance chip above.