Pith. sign in

REVIEW 4 major objections 5 minor 37 references

ROSA: Addressing text understanding challenges in photographs via ROtated SAmpling

T0 review · 4 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read ROSA, an inference-time strategy that samples answers across four rotated copies of an image, lifts exact-match VQA accuracy by up to 11.7 absolute points on photos with misoriented text.

desk verdict A simple, model-agnostic inference-time rotation-and-sampling trick that consistently improves VQA on misoriented text, but the headline number rests on an evaluation split that mixes natural and synthetic cases and needs clearer reporting. read the letter →

arxiv 2506.03665 v1 pith:4SFPQO2S submitted 2025-06-04 cs.CL cs.CV

Hernán Maina , Guido Ivetta , Mateo Lione Stuto , Julian Martin Eisenschlos , Jorge Sánchez , Luciana Benotti This is my paper
classification cs.CLcs.CV
keywords visualquestionansweringrotatedtextdecodingstrategyVizWiz-VQAimpairmentmultimodalmodelsinference-timeaugmentationreadinginimages
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

ROSA is an inference-time decoding strategy for visual question answering (VQA) aimed at photographs whose text is not upright. The paper argues that presenting the same image to a multimodal model in four canonical rotations (0°, 90°, 180°, 270°), sampling several candidate answers per rotation, and returning the candidate with the highest token likelihood substantially improves correctness when the answer requires reading misoriented text. The headline result is an 11.7 absolute exact-match gain over greedy decoding on the VizWiz(Conventional) subset, with consistent gains across every model tested on misoriented-text datasets. The motivation comes from interviews with eight visually impaired participants, whose framing habits often leave text rotated, and the paper builds evaluation subsets from VizWiz-VQA to reflect those habits. Because ROSA needs no retraining, it is a drop-in wrapper that could improve assistive VQA systems where misoriented photos are common.

What carries the argument

The load-bearing mechanism is a decoding rule that combines image rotation, stochastic sampling, and likelihood selection. For each of four rotations (0°, 90°, 180°, 270°), the model samples candidate answers with temperature 0.5, top_p 1.0, and top_k 50; every candidate is scored by its joint token probability $p(a|v,q) = \prod_{i=1}^{n} p(t_i \mid t_1, \ldots, t_{i-1}, v, q)$, and the answer with the highest score is returned, with majority voting to break ties. The four angles come from an orientation analysis of the VizWiz-Text images performed with a text-detection model, which found landscape and portrait orientations dominant and justified assuming the symmetric rotations are present. The paper isolates the mechanism's active ingredient with two ablations: RepeatRot uses the same sampling but no rotation, and RandomRot uses random angles; both underperform full ROSA on misoriented-text datasets, indicating that the structured rotations, not sampling alone or arbitrary augmentation, drive the improvement.

What would settle it

Build or select a set of VQA images whose text sits at intermediate angles such as 45°, 135°, or 225°, or whose text regions have different orientations in one image, run ROSA and greedy decoding with the same models, and compare exact-match accuracy; if ROSA's advantage disappears or reverses, the four-canonical-angle assumption is the limiting factor. A second check is to compare the rotation that yields the highest-likelihood answer against the angle a text detector identifies as upright on images with known text orientation.

Watch

Extended reading notes

Core claim

The paper's central discovery is that current multimodal models lose about 7.7 points of exact-match accuracy on average when the text they must read is rotated, and that this weakness can be reduced at inference time, with no retraining, by scoring sampled answers across four rotated versions of the image. The selection rule uses $p(a|v,q) = \prod_{i=1}^{n} p(t_i \mid t_1, \ldots, t_{i-1}, v, q)$ to score each candidate answer and returns the highest-likelihood one. In the VizWiz(Conventional) subset, built from photos taken by visually impaired people, ROSA raises exact match from 45.1 to 56.8 for the best model, and the median gain over greedy decoding across all tested models is 10.3 absolute points. Controlled baselines show the gain comes from the fixed rotation grid: RepeatRot (sampling without rotations) and RandomRot (random angles) both trail ROSA. On well-oriented text, greedy decoding generally remains better, so ROSA is a targeted fix for misoriented text rather than a general replacement.

Load-bearing premise

The load-bearing premise is that the text needed for the answer has one dominant orientation that one of the four right-angle rotations can bring upright, and that token-likelihood scores stay comparable across rotations so the highest-scoring answer is the most reliable; the paper's own Figure 1 shows photos containing multiple text orientations that no single rotation can fix.

Editorial extensions

If this is right

  • Adopting ROSA in VQA pipelines that serve visually impaired users should improve answers on photos with rotated text by several absolute points, at the cost of multiple inferences per question.
  • Because ROSA is inference-only, it can be layered on any current or future black-box multimodal model without retraining or fine-tuning.
  • On datasets where text is already upright, greedy decoding generally matches or beats ROSA, so a practical system should apply ROSA selectively rather than universally.
  • The observed gains on artificially rotated OCR-VQA suggest the benefit generalizes beyond VizWiz's realistic photos to any text-bearing image with non-canonical orientation.
  • Structured 90-degree rotations outperform random rotations, which indicates the canonical grid, not augmentation diversity, is what carries the improvement.

Reading between the lines

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

  • If the single-orientation premise fails for images with text at intermediate angles or in multiple orientations, a case the paper's own Figure 1 shows, ROSA's advantage should shrink; a natural extension is finer rotation grids or rotating detected text regions individually.
  • The paper does not test whether token likelihoods are calibrated across rotations; if they are not, a normalized or ensemble-aggregated score could be more reliable than raw highest likelihood.
  • ROSA's idea transfers naturally to other text-reading tasks with orientation variability, such as document photo OCR or packaging recognition, and could be combined with abstention mechanisms so the model declines to answer when no rotation yields a confident response.
  • A testable practical extension is an adaptive selector that runs ROSA only when an orientation detector flags misoriented text, saving compute on the well-oriented majority while preserving gains on the rotated minority.
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 addresses the problem of Visual Question Answering (VQA) on photographs, taken mainly by visually impaired users, in which the text needed to answer the question is misoriented. The authors first report interviews with visually impaired participants to identify common framing and orientation conventions, and use these insights to construct evaluation subsets from VizWiz-VQA, including a mixed subset called VizWiz(Conventional) that contains both naturally misoriented images and synthetic rotations of well-oriented images. They then propose ROSA (ROtated SAmpling), an inference-time decoding strategy that presents the model with the same image rotated by 0, 90, 180, and 270 degrees, samples multiple answers per rotation, and returns the sampled answer with the highest sequence likelihood. Across six large multimodal models and three metrics, ROSA is reported to outperform Greedy decoding on all misoriented-text datasets, with the largest gain being 11.7 absolute points in exact-match accuracy on VizWiz(Conventional) for Qwen2.5-VL-3B. Two controlled baselines, RepeatRot and RandomRot, are used to separate the effects of sampling and of the particular rotation set. The paper concludes that ROSA is a simple, retraining-free way to improve robustness to misoriented text in VQA.

Significance. If the reported results hold, ROSA is a practically useful and refreshingly simple inference-time method: it requires no retraining, is model-agnostic, and consistently improves Greedy decoding across several recent multimodal models and three evaluation metrics. The inclusion of RepeatRot and RandomRot baselines is a genuine strength, because it isolates the contribution of the structured rotations from the contribution of sampling alone and from arbitrary rotations. The community-informed evaluation subsets and the interview findings are also valuable contributions to the study of assistive VQA. However, the strength of the central empirical claim is currently limited by the construction of the main evaluation subset and by the overlap between the rotation-set design and the evaluation pool; these issues are described in the major comments.

major comments (4)
  1. [Section 4.1.1, Table 1] VizWiz(Conventional), the subset on which the headline 11.7-point gain is reported, is a mixture of two very different kinds of examples: (1) genuine VizWiz images whose text is rotated by at least 90 degrees, and (2) synthetic 90/180/270-degree rotations of images that originally had well-oriented text. The paper never reports results separately for these two subgroups. On the synthetic portion, ROSA only needs to discover the single applied rotation to restore a clean, otherwise easy image, so gains on that portion may not transfer to the authentic multi-orientation, blurry, and partially framed photographs shown in Figure 1. Please report EM, Acc, and SAS separately for the genuine misoriented images and for the synthetic rotated variants, with the number of samples in each subgroup, so that the real-world claim can be evaluated.
  2. [Section 5, Figure 2] The four ROSA rotations were selected based on 'an empirical analysis conducted on the 1270 images that make up the VizWiz-Text dataset,' and the evaluation subsets are drawn from this same VizWiz-Text pool. This design/evaluation overlap means the rotation set is effectively tuned to the test distribution. In addition, the text detector used for the orientation analysis could not distinguish symmetric rotations, so the 90-degree and 180-degree rotations are assumed rather than measured despite Figure 2 showing no detections at those angles. Please validate the rotation set on a held-out set (for example, unused VizWiz partitions or a different misoriented-text corpus) and/or ablate the rotation set to show that the choice of 0/90/180/270 is not overfit to the evaluation distribution.
  3. [Section 5, 'Inference process'] The sampling protocol is incompletely specified. The text says that 'the model generates multiple independent predictions' for each rotation but never states the number of samples per rotation or the total number of inferences per image. This number is essential for reproducibility, for interpreting the computational overhead that the paper itself acknowledges as a limitation, and for ensuring that the RepeatRot and RandomRot baselines use the same inference budget as ROSA. Please specify the exact sampling budget and state whether it was held constant across models and datasets.
  4. [Section 6.2, Table 1] Most numbers in Table 1 are single-run results without variance estimates or significance tests; only a few values are marked as averaged over five runs. Several differences between Greedy and ROSA are small or negative (for example, Molmo-7B on OCR-VQA(Oriented) EM 42.0 vs. 41.8, or GPT-4o-mini on OCR-VQA(Oriented) SAS 79.9 vs. 79.2), so it is unclear how much of the reported consistency is robust. Please provide repeated-run means with standard deviations or confidence intervals, and/or significance tests, for at least the headline comparisons in Table 1.
minor comments (5)
  1. [Section 4] There are several typos and spacing errors, including 'aplplied', 'datsets', 'VizWiz-VQAconsists', 'theVizWiz', and 'theZeng'; these should be corrected.
  2. [Section 5, tie-breaking] The phrase 'In case of ties among the top k predictions' is ambiguous: it is unclear whether k refers to the top_k sampling parameter, the number of sampled answers, or something else. Please define the tie-breaking rule precisely.
  3. [Section 6.2, oracle discussion] The 'oracle' is defined as Greedy decoding on OCR-VQA(Oriented), which is a different dataset from VizWiz(Conventional). This is a useful reference point, but it is not an upper bound on the same task, and the text should say so explicitly rather than implying that ROSA is close to the ceiling for the VizWiz-based evaluation.
  4. [Figure 3] The caption should explain how 'contribution' is computed and why combinations below 3% are excluded; currently the reader must infer these details from the body text.
  5. [Abstract and Table 1] The abstract reports '11.7 absolute points' without naming the model or dataset; please identify the setting (Qwen2.5-VL-3B on VizWiz(Conventional)) to avoid overgeneralizing the headline number.

Circularity Check

1 steps flagged · score 6.0 of 10

ROSA's headline gain is partly built into the evaluation design: the rotation set is selected from the same VizWiz-Text pool used to construct the test subsets, whose synthetic half is generated by applying exactly those rotations to clean images.

  1. fitted input called prediction [Section 4.1.1 (Derived datasets) and Section 5 (Motivation and rationale)]
    "VizWiz (Conventional) ... includes 990 samples from VizWiz-Text, comprising: (1) all pairs with images containing text rotated ≥ |90|◦, and (2) specific rotated variants (at 90°, 180°, and/or 270°) of images with originally well-oriented text. ... To implement this approach, we selected four key rotations: 0°, 90°, 180°, and 270°. This decision was based on an empirical analysis conducted on the 1270 images that make up the VizWiz-Text dataset."

    The only tuned design choice of ROSA (the rotation set) is selected from the orientation statistics of the same VizWiz-Text pool from which every evaluation subset is drawn, and the synthetic component of VizWiz(Conventional) is generated by applying exactly those same 90°/180°/270° rotations to clean, well-oriented images. On that synthetic half, one of ROSA's four candidates is, by construction, the original unrotated image that Greedy never sees; so the headline 11.7-point EM improvement partly measures the method's ability to recover an orientation that the evaluators inserted, rather than robustness to naturally misoriented photographs. The paper does not report the natural-photo and synthetic-rotation subgroups separately, so the real-world claim is not independently established.

full rationale

The likelihood-based selection in Section 5 is a standard chain-rule computation with no fitted constant, and the comparison against RepeatRot and RandomRot gives the central claim some independent empirical content. However, the paper's evaluation design creates a partial circularity: the four ROSA rotations were chosen from an empirical analysis of the same 1270-image VizWiz-Text pool used to build the test subsets, and VizWiz(Conventional) includes synthetic rotations at exactly those angles applied to clean, well-oriented images. Thus the favorable condition for ROSA is partially constructed into the test set rather than predicted from first principles. The self-citation to Eisenschlos et al. (2024) in the Limitations is not load-bearing. Overall, the central claim has independent content but the headline comparison is partially circular by construction.

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

The method depends on a small set of chosen hyperparameters (temperature, top_k, top_p, rotation set), plus an unstated sample count, and on several domain assumptions about misorientation structure and likelihood calibration. No new entities are introduced.

free parameters (5)
  • temperature = 0.5
    Chosen to balance sampling variability and coherence (Section 5); no ablation shown.
  • top_k = 50
    Chosen for sampling (Section 5); no ablation shown.
  • top_p = 1.0
    Set to 1.0, effectively disabling nucleus filtering (Section 5).
  • number of samples per rotation = not stated
    The paper says 'multiple independent predictions' per rotation but never specifies the count; this is required to reproduce ROSA's reported numbers.
  • rotation set = {0°, 90°, 180°, 270°}
    Chosen based on an empirical orientation analysis of the evaluation dataset (Section 5); this is test-distribution-informed hyperparameter choice.
assumptions (4)
  • domain assumption The relevant text in each image has a single global orientation that one of the four canonical rotations can correct.
    Required for ROSA to be effective; Section 5 motivation and Figure 1 show multi-orientation cases where this fails, but no failure-rate analysis is provided.
  • domain assumption Model token likelihoods (p(a|v,q)) are comparable across rotated inputs and correlate with answer correctness.
    ROSA selects the highest-likelihood answer across rotations; if likelihoods are miscalibrated across orientations, selection could be biased (Section 5).
  • domain assumption The constructed evaluation subsets (VizWiz-Conventional, VizWiz-Random) represent real-world misoriented text distributions.
    The subsets include synthetic rotations of well-oriented images (Section 4.1.1), which may not match naturally misoriented photos.
  • ad hoc to paper The orientation statistics from the detection model used in Section 5 are reliable despite the model's inability to resolve symmetric rotations.
    The paper acknowledges the detector reported no text at 90° or 180°, yet assumes such orientations exist and chooses all four rotations anyway.

how reviews work

0 comments
Cite this review

Pith. "Pith review of ROSA: Addressing text understanding challenges in photographs via ROtated SAmpling." pith.science (2026). https://pith.science/paper/4SFPQO2S

@misc{pith2026250603665,
  author       = {Pith},
  title        = {Pith review of: ROSA: Addressing text understanding challenges in photographs via ROtated SAmpling},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/4SFPQO2S}},
  note         = {Machine review of arXiv:2506.03665}
}
read the original abstract

Visually impaired people could benefit from Visual Question Answering (VQA) systems to interpret text in their surroundings. However, current models often struggle with recognizing text in the photos taken by this population. Through in-depth interviews with visually impaired individuals, we identified common framing conventions that frequently result in misaligned text. Existing VQA benchmarks primarily feature well-oriented text captured by sighted users, under-representing these challenges. To address this gap, we introduce ROtated SAmpling (ROSA), a decoding strategy that enhances VQA performance in text-rich images with incorrectly oriented text. ROSA outperforms Greedy decoding by 11.7 absolute points in the best-performing model.

Figures

Figures reproduced from arXiv: 2506.03665 by the authors.

Figure 1
Figure 1. Examples of image pairs and associated vi [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Polar histogram of text orientations in the [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗
Figure 3
Figure 3. Distribution of final correct EM-based pre [PITH_FULL_IMAGE:figures/full_fig_p010_3.png] view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

37 extracted references · 20 canonical work pages

  1. [1]

    URL: " 'urlintro :=

    ENTRY address author booktitle chapter edition editor howpublished institution journal key month note number organization pages publisher school series title type volume year eprint doi pubmed url lastchecked label extra.label sort.label short.list INTEGERS output.state before.all mid.sentence after.sentence after.block STRINGS urlintro eprinturl eprintpr...

  2. [2]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 global.max substring 't := if while FUNCTION word.in bbl.in capitalize " " * FUNCT...

  3. [3]

    Lawrence Zitnick, and Devi Parikh

    Stanislaw Antol, Aishwarya Agrawal, Jiasen Lu, Margaret Mitchell, Dhruv Batra, C. Lawrence Zitnick, and Devi Parikh. 2015. https://openaccess.thecvf.com/content_iccv_2015/html/Antol_VQA_Visual_Question_ICCV_2015_paper.html VQA: Visual Question Answering . In Proceedings of the IEEE International Conference on Computer Vision (ICCV)

    work page 2015

  4. [4]

    Improving Text Proposals for Scene Images with Fully Convolutional Networks

    Dena Bazazian, Raul Gomez, Anguelos Nicolaou, Lluis Gomez, Dimosthenis Karatzas, and Andrew D. Bagdanov. 2017. http://arxiv.org/abs/1702.05089 Improving Text Proposals for Scene Images with Fully Convolutional Networks

    work page Pith review arXiv 2017

  5. [5]

    Lucas Beyer, Andreas Steiner, André Susano Pinto, Alexander Kolesnikov, Xiao Wang, Daniel Salz, Maxim Neumann, Ibrahim Alabdulmohsin, Michael Tschannen, Emanuele Bugliarello, Thomas Unterthiner, Daniel Keysers, Skanda Koppula, Fangyu Liu, Adam Grycner, Alexey Gritsenko, Neil Houlsby, Manoj Kumar, Keran Rong, Julian Eisenschlos, Rishabh Kabra, Matthias Bau...

    arXiv 2024

  6. [6]

    Bigham, Chandrika Jayant, Hanjie Ji, Greg Little, Andrew Miller, Robert C

    Jeffrey P. Bigham, Chandrika Jayant, Hanjie Ji, Greg Little, Andrew Miller, Robert C. Miller, Robin Miller, Aubrey Tatarowicz, Brandyn White, Samual White, and Tom Yeh. 2010. https://doi.org/10.1145/1866029.1866080 V iz W iz: nearly real-time answers to visual questions . In Proceedings of the 23nd Annual ACM Symposium on User Interface Software and Techn...

    arXiv 2010

  7. [7]

    Jawahar, Ernest Valveny, and Dimosthenis Karatzas

    Ali Furkan Biten, Rubèn Tito, Andrés Mafla, Lluis Gomez, Marçal Rusiñol, C.V. Jawahar, Ernest Valveny, and Dimosthenis Karatzas. 2019. https://doi.org/10.1109/ICCV.2019.00439 Scene Text Visual Question Answering . In 2019 IEEE/CVF International Conference on Computer Vision (ICCV), pages 4290--4300

    arXiv 2019

  8. [8]

    Smith, Hannaneh Hajishirzi, Ross Girshick, Ali Farhadi, and Aniruddha Kembhavi

    Matt Deitke, Christopher Clark, Sangho Lee, Rohun Tripathi, Yue Yang, Jae Sung Park, Mohammadreza Salehi, Niklas Muennighoff, Kyle Lo, Luca Soldaini, Jiasen Lu, Taira Anderson, Erin Bransom, Kiana Ehsani, Huong Ngo, YenSung Chen, Ajay Patel, Mark Yatskar, Chris Callison-Burch, Andrew Head, Rose Hendrix, Favyen Bastani, Eli VanderBilt, Nathan Lambert, Yvon...

    arXiv 2024

Show all 37 references
  1. [9]

    Julian Eisenschlos, Hern \'a n Maina, Guido Ivetta, and Luciana Benotti. 2024. https://doi.org/10.18653/v1/2024.findings-acl.250 Selectively answering visual questions . In Findings of the Association for Computational Linguistics: ACL 2024, pages 4219--4229, Bangkok, Thailand...

    2024 doi

  2. [10]

    Logan Engstrom, Brandon Tran, Dimitris Tsipras, Ludwig Schmidt, and Aleksander Madry. 2019. https://openreview.net/forum?id=BJfvknCqFQ A rotation and a translation suffice: Fooling CNN s with simple transformations

    2019

  3. [11]

    Yarin Gal and Zoubin Ghahramani. 2016. https://proceedings.mlr.press/v48/gal16.html Dropout as a bayesian approximation: Representing model uncertainty in deep learning . In Proceedings of The 33rd International Conference on Machine Learning, volume 48 of Proceedings of Machi...

    2016

  4. [12]

    Stangl, Anhong Guo, Chi Lin, Kristen Grauman, Jiebo Luo, and Jeffrey P

    Danna Gurari, Qing Li, Abigale J. Stangl, Anhong Guo, Chi Lin, Kristen Grauman, Jiebo Luo, and Jeffrey P. Bigham. 2018. https://doi.org/10.1109/CVPR.2018.00380 V iz W iz G rand C hallenge: A nswering V isual Q uestions from B lind P eople . In 2018 IEEE/CVF Conference on Compu...

    2018

  5. [13]

    Tong He, Weilin Huang, Yu Qiao, and Jian Yao. 2016. http://arxiv.org/abs/1603.09423 Accurate Text Localization in Natural Image with Cascaded Convolutional Text Network

    2016 arXiv

  6. [14]

    Wenhao He, Xu-Yao Zhang, Fei Yin, and Cheng-Lin Liu. 2018. https://doi.org/10.1109/TIP.2018.2855399 Multi-Oriented and Multi-Lingual Scene Text Detection With Direct Regression . IEEE Transactions on Image Processing, 27(11):5406--5419

    2018

  7. [15]

    Ronghang Hu, Amanpreet Singh, Trevor Darrell, and Marcus Rohrbach. 2020. https://doi.org/10.1109/CVPR42600.2020.01001 Iterative Answer Prediction With Pointer-Augmented Multimodal Transformers for TextVQA . In 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition...

    2020

  8. [16]

    Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. 2012. https://proceedings.neurips.cc/paper_files/paper/2012/file/c399862d3b9d6b76c8436e924a68c45b-Paper.pdf Imagenet classification with deep convolutional neural networks . In Advances in Neural Information Processing Sy...

    2012

  9. [17]

    Kenton Lee, Mandar Joshi, Iulia Turc, Hexiang Hu, Fangyu Liu, Julian Eisenschlos, Urvashi Khandelwal, Peter Shaw, Ming-Wei Chang, and Kristina Toutanova. 2023. http://arxiv.org/abs/2210.03347 Pix2struct: Screenshot parsing as pretraining for visual language understanding

    2023 arXiv

  10. [18]

    David Kellogg Lewis. 1969. Convention: A Philosophical Study. Wiley-Blackwell, Cambridge, MA, USA

    1969

  11. [19]

    Minghui Liao, Zhen Zhu, Baoguang Shi, Gui-song Xia, and Xiang Bai. 2018. https://doi.org/10.1109/CVPR.2018.00619 Rotation-sensitive regression for oriented scene text detection . In 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5909--5918

    2018

  12. [20]

    Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. 2024 a . https://doi.org/10.1109/CVPR52733.2024.02484 Improved baselines with visual instruction tuning . In 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 26296--26306

    2024

  13. [21]

    Haotian Liu, Chunyuan Li, Yuheng Li, Bo Li, Yuanhan Zhang, Sheng Shen, and Yong Jae Lee. 2024 b . https://llava-vl.github.io/blog/2024-01-30-llava-next/ LLaVA-NeXT: Improved reasoning, OCR, and world knowledge

    2024

  14. [22]

    Libuše Ludíková and Dita Finková. 2012. https://doi.org/https://doi.org/10.1016/j.sbspro.2012.09.587 Improvement in Education of People with Visual Impairment . Procedia - Social and Behavioral Sciences, 55:971--979. 3rd. International Conference on New Horizons in Education -...

    2012 doi

  15. [23]

    Jianqi Ma, Weiyuan Shao, Hao Ye, Li Wang, Hong Wang, Yingbin Zheng, and Xiangyang Xue. 2018. https://doi.org/10.1109/TMM.2018.2818020 Arbitrary-Oriented Scene Text Detection via Rotation Proposals . IEEE Transactions on Multimedia, 20(11):3111--3122

    2018

  16. [24]

    Hana Majerova. 2017. https://doi.org/https://doi.org/10.1016/j.sbspro.2017.02.117 The Person in a Situation of Visual Impairment and its Perception and Imagination from the Qualitative Viewpoint . Procedia - Social and Behavioral Sciences, 237:751--757. Education, Health and I...

    2017 doi

  17. [25]

    Meta AI . 2024. http://arxiv.org/abs/2407.21783 The llama 3 herd of models

    2024 arXiv

  18. [26]

    Anand Mishra, Shashank Shekhar, Ajeet Kumar Singh, and Anirban Chakraborty. 2019. https://doi.org/10.1109/ICDAR.2019.00156 OCR-VQA: Visual Question Answering by Reading Text in Images . In 2019 International Conference on Document Analysis and Recognition (ICDAR), pages 947--952

    2019

  19. [27]

    OpenAI. 2024. http://arxiv.org/abs/2303.08774 Gpt-4 technical report

    2024 arXiv

  20. [28]

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

    2025

  21. [29]

    Connor Shorten and Taghi M Khoshgoftaar. 2019. https://journalofbigdata.springeropen.com/articles/10.1186/s40537-019-0197-0#citeas A survey on image data augmentation for deep learning . Journal of Big Data, 6(1):1--48

    2019 doi

  22. [30]

    Karen Simonyan and Andrew Zisserman. 2015. https://www.robots.ox.ac.uk/ vgg/publications/2015/Simonyan15/ Very deep convolutional networks for large-scale image recognition . In International Conference on Learning Representations

    2015

  23. [31]

    Amanpreet Singh, Vivek Natarajan, Meet Shah, Yu Jiang, Xinlei Chen, Dhruv Batra, Devi Parikh, and Marcus Rohrbach. 2019. https://doi.org/10.1109/CVPR.2019.00851 Towards VQA Models That Can Read . In 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pa...

    2019

  24. [32]

    Wenhui Wang, Furu Wei, Li Dong, Hangbo Bao, Nan Yang, and Ming Zhou. 2020. https://proceedings.neurips.cc/paper/2020/hash/3f5ee243547dee91fbd053c1c4a845aa-Abstract.html MiniLM: Deep Self-Attention Distillation for Task-Agnostic Compression of Pre-Trained Transformers . In Proc...

    2020

  25. [33]

    Chi, Sharan Narang, Aakanksha Chowdhery, and Denny Zhou

    Xuezhi Wang, Jason Wei, Dale Schuurmans, Quoc V Le, Ed H. Chi, Sharan Narang, Aakanksha Chowdhery, and Denny Zhou. 2023. https://openreview.net/forum?id=1PL1NIMMrw Self-consistency improves chain of thought reasoning in language models . In The Eleventh International Conferenc...

    2023

  26. [34]

    Zhengyuan Yang, Yijuan Lu, Jianfeng Wang, Xi Yin, Dinei Florencio, Lijuan Wang, Cha Zhang, Lei Zhang, and Jiebo Luo. 2021. Tap: Text-aware pre-training for text-vqa and text-caption. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), p...

    2021

  27. [35]

    Cong Yao, Xiang Bai, Nong Sang, Xinyu Zhou, Shuchang Zhou, and Zhimin Cao. 2016. http://arxiv.org/abs/1606.09002 Scene Text Detection via Holistic, Multi-Channel Prediction

    2016 arXiv

  28. [36]

    Xiaoyu Zeng, Yanan Wang, Tai-Yin Chiu, Nilavra Bhattacharya, and Danna Gurari. 2020. https://doi.org/10.1145/3415220 Vision Skills Needed to Answer Visual Questions . Proceedings of the Association for Computing Machinery (ACM) on Human-Computer Interaction, 4(CSCW2)

    2020 doi

  29. [37]

    Shi-Xue Zhang, Chun Yang, Xiaobin Zhu, and Xu-Cheng Yin. 2024. https://doi.org/10.1109/TMM.2023.3286657 Arbitrary Shape Text Detection via Boundary Transformer . IEEE Transactions on Multimedia, 26:1747--1760

    2024

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.