Hear the Scene: Audio-Enhanced Text Spotting

Bo Wang; Jing Li

REVIEW 5 major objections 5 minor 14 references

Hear the Scene: Audio-Enhanced Text Spotting

T0 review · 5 major / 5 minor · reviewed 2026-08-11 · deepseek-v4-flash

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

Pith's one-line read Scene text spotting can be trained on transcriptions alone: the model's recognition attention map doubles as its location signal, matching fully supervised performance closely.

desk verdict A transcription-only spotter with a misleading headline metric; the core idea is already in Tang et al. 2022, and the reported gains vanish under the standard polygon metric. read the letter →

arxiv 2412.19504 v3 pith:DQ7WGH2J submitted 2024-12-27 cs.CV

Jing Li , Bo Wang This is my paper · ORCID

classification cs.CV

keywords scenetextspottingtranscription-onlysupervisionweaklysupervisedlocalizationquery-basedtransformerattentionactivationmapcircularcurriculumlearningaudioannotationend-to-endrecognition

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 scene text spotting—locating and reading words in photos—does not need geometric annotations such as bounding boxes, polygons, or even single points. The proposed model, EchoSpot, trains on nothing but the transcription of the text, plus the image. Location is recovered implicitly from the interaction of text queries with image embeddings and then sharpened by the attention activation map produced during recognition. The authors report that this transcription-only training lands within a few points of fully supervised baselines on public benchmarks and stays competitive with a single-point-supervised method, while also allowing annotations to be spoken aloud rather than typed. If the claim holds, the cost of building text-spotting datasets drops sharply and the labeling process becomes accessible to people who cannot use a keyboard or mouse.

What carries the argument

The mechanism doing the work is the coarse-to-fine cross-attention localization pipeline. A set of text queries first attends to image embeddings to produce a coarse attention activation map; that map is converted into a binary attention mask, and a second transformer decoder refines the mask together with the image embeddings into a sharper text location. Training is driven by a text-based Hungarian matching loss that aligns predicted sequences to ground-truth transcriptions, so error gradients flow from recognition back through the attention maps with no geometric supervision. The circular curriculum learning strategy schedules training from simple to complex text images and is the component credited with making from-scratch weak supervision converge.

What would settle it

Measure the overlap (IoU) between the final coarse-to-fine attention masks and ground-truth text polygons on a held-out set. If mask IoU stays at chance level while end-to-end recognition accuracy is close to fully supervised, then the reported accuracy is not being carried by the attention localization mechanism as claimed; conversely, if mask IoU tracks recognition accuracy, the attention proxy is doing the work.

Watch

Extended reading notes

Core claim

EchoSpot's central proposal is that text location is a byproduct of text recognition. With only ground-truth transcriptions, randomly initialized text queries interact with image embeddings through cross-attention, and the resulting attention activation map—normally an internal artifact of recognition—is treated as a coarse signal for where the text is. A second cross-attention stage refines this coarse map into a final mask, and a text-based Hungarian matching loss lets recognition gradients supervise the entire localization pathway without any geometric loss. A circular curriculum, which starts on easy text instances and gradually introduces harder scenes, is introduced specifically to make this weakly supervised pipeline converge from scratch. The experiments on four benchmarks place EchoSpot within a few points of fully supervised spotters and close to the single-point-supervised baseline, which the paper reads as evidence that location annotations can be dropped without sacrificing competitive accuracy.

Load-bearing premise

The load-bearing premise is that the attention activation map generated while the model recognizes text is a dependable enough stand-in for text location that the model can learn to localize from it, even though no geometric label is ever shown.

Editorial extensions

If this is right

Scene text spotting datasets could be annotated with plain text transcriptions only, eliminating the need for bounding boxes, polygons, or point labels.
Because transcriptions can be provided by voice and turned into text by speech recognition, annotation becomes faster and opens dataset creation to visually impaired contributors; the reported audio-based results on Total-Text stay within about one point of typed transcription.
A transcription-only spotter can reach competitive end-to-end accuracy on both horizontal benchmarks (ICDAR 2013/2015) and curved-text benchmarks (Total-Text, SCUT-CTW1500), narrowing the gap to fully supervised methods.
The ablation indicates that both the coarse-to-fine localization mechanism and the circular curriculum are needed: removing either lowers end-to-end performance or slows convergence.
By treating spotting as a captioning-style task, the framework lets advances in vision-language attention transfer directly to text spotting.

Reading between the lines

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

If the attention-activation proxy is as reliable as reported, the learned masks could be distilled into a lightweight detection head, yielding box or polygon outputs with no extra geometric labels—something the paper does not test.
The same transcription-only recipe could be applied to video text spotting, where a single spoken transcription could supervise many frames; whether temporal continuity helps or hurts the implicit attention proxy is an open question.
Audio annotation could be combined with crowdsourced audio descriptions or accessibility metadata, turning speech intended for human users into training signal for spotting models.
A decisive test of the mechanism would be a head-to-head comparison with the earlier transcription-only, voice-annotated method cited in the paper, using the same backbone and benchmarks, to isolate which components actually close the gap to full supervision.

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.

Desk Editor's Note

read the letter

Colleague,

Two things before you spend time on this. The headline claim that transcription-only training can produce competitive text spotting is not new—Tang et al. (2022) already demonstrated it with voice annotations, and this paper cites them but never compares against them. Worse, the 'competitive' claim only holds under the single-point metric. Under the standard polygon metric, EchoSpot is 10–25 points behind the fully-supervised baselines in every table (e.g., 77.7 vs 88.8 on IC13; 60.2 vs 83.3 on IC15). The abstract does not qualify the claim, so it is misleading.

What is actually decent: the architecture is a sensible combination of query-based cross-attention, an attention-activation mask for localization, and a coarse-to-fine refinement. The circular curriculum makes sense for weakly supervised training. The ablation (Table 1) shows that both the coarse-to-fine mechanism and the curriculum contribute, and the audio annotation experiment (Table 6) shows only a small drop from text transcription, which is a practical plus. The single-point numbers are close to SPTS, so the method is not a dud.

The soft spots are more than cosmetic. First, the metric mismatch is load-bearing: the paper alternates between single-point and polygon without noting that fully-supervised baselines are evaluated only under the polygon metric. That makes the central comparison invalid. Second, no comparison with Tang et al. 2022, the obvious baseline for the claimed contribution. Third, Table 1 shows 2 coarse-to-fine iterations outperform 1, yet the main tables report results that match the 1-iteration row (e.g., Total-Text None 65.1). The paper never explains why. Fourth, no error bars anywhere; with small benchmark differences, that matters. Finally, the writing is rough: missing references in the related work, broken citations, and a conclusion that repeats the unsupported 'competitive' claim.

If you read it for the architecture details, there's some value there. As a submission, though, the central claim is not supported by the data as reported. I would desk-reject it. If the authors fix the metric reporting, add the Tang et al. comparison, and resolve the iteration inconsistency, it could become a workshop-level contribution. Not more.

Recommendation: do not send to peer review.

Referee Report

5 major / 5 minor

Summary. The paper proposes EchoSpot, a query-based transformer for scene text spotting trained from transcription-only annotations. The method uses coarse-to-fine cross-attention, a Hungarian matching loss, circular curriculum learning, and an audio-annotation interface. Experiments are reported on ICDAR 2013, ICDAR 2015, Total-Text, and SCUT-CTW1500 under both a single-point and a polygon evaluation metric. The central claim is that competitive text-spotting accuracy can be achieved without location annotations, reducing annotation cost and improving accessibility.

Significance. If the results were fully validated, EchoSpot would be a meaningful step toward reducing text-spotting annotation costs and making annotation accessible to a broader set of contributors. The core idea of inferring location from recognition-stage attention, combined with curriculum learning, is interesting and worth exploring. However, the current evidence does not support the headline claim: the polygon-metric results in the paper's own tables are far below all fully-supervised baselines, the most relevant prior work on transcription-only and audio annotation (Tang et al. 2022b) is never compared, and the ablation is internally inconsistent. The paper also provides no code or error bars, so the reported numbers cannot be independently verified.

major comments (5)

[Tables 2–5, Sec. 4.2] The headline claim of competitive performance is contradicted by the paper's own polygon-metric results. For example, on ICDAR 2015 the EchoSpot polygon row reports 60.2, while Mask TextSpotter v3 reports 83.3; on Total-Text the polygon row reports 61.5, while MANGO reports 72.9. The fully-supervised baselines are evaluated with the standard polygon end-to-end metric, so reporting EchoSpot under the single-point metric in the same comparison is not apples-to-apples. The abstract and conclusion do not state that 'competitive' refers only to the relaxed single-point metric. The authors should either report all methods under the same metric or provide a clear justification for why the single-point metric is the appropriate basis for comparison; under the standard metric, the central claim is not supported.
[Sec. 2.2, Ref. Tang et al. [2022b]] Tang et al. (2022b), 'You can even annotate text with voice: Transcription-only-supervised text spotting,' which addresses exactly the same transcription-only and audio-annotation setting, is cited in the reference list but never compared against in Tables 2–6. Without a direct comparison to this most relevant baseline, the claim that EchoSpot provides a competitive transcription-only approach is unsubstantiated. The authors should add this baseline or state explicitly why comparison is not possible, and discuss any differences in evaluation setup.
[Table 1, Tables 4–5] The ablation in Table 1 shows that two coarse-to-fine iterations outperform one (75.2 vs. 74.8 under the full lexicon on Total-Text), yet the main results in Table 4 (65.1/74.8) match the single-iteration row. The paper never states which configuration is used in the main tables or why the configuration with the best ablation result was not adopted. This inconsistency makes the reported numbers hard to interpret and requires clarification.
[Sec. 3.2] The 'Text-based Hungarian Matching Loss' is described only in prose. No equation, pseudocode, or precise definition is given, so the reader cannot verify how the Hungarian algorithm is applied to predicted and ground-truth text sequences, or how gradients are backpropagated to the localization branch. This is a load-bearing component because the entire method rests on learning locations from transcription-only supervision through this loss.
[Sec. 3.4, Table 6] The audio-annotation claim is not experimentally supported. Table 6 reports results for 'Audio (Word)' and 'Audio (Character)' annotation styles, but Section 3.4 provides no details on the speech recognition pipeline, no annotation-time measurements, and no comparison with manual transcription time. The statement that audio annotation 'significantly diminishes annotation time' is therefore not evidenced by the presented experiments.

minor comments (5)

[Sec. 2.1] The citations 'Li et al. Shi et al. [2016]' and 'He et al. Xing et al. [2019]' have incorrect author names; they should be Shi et al. and Xing et al., respectively.
[Introduction, Tables 4–5] The introduction contains the malformed citation 'Zhao et al. [2024b], ?]', and Tables 4 and 5 list 'ABCNet v2 ?' without a proper reference. These should be fixed.
[Sec. 3.2] The subsection heading 'Text-based Hungarian Matching Loss' is not formatted as a separate heading and appears to run into the preceding paragraph; it should be given its own subsection.
[Sec. 4.3] Implementation details lack the training schedule, batch size, input resolution, number of epochs, and ASR configuration; these are needed for reproducibility.
[Tables 2–5] The table captions should state explicitly which metric is used for each row group; the row labels 'Single-Point' and 'Polygon' are ambiguous when compared against the fully-supervised rows.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: EchoSpot is an empirical system evaluated on external benchmarks; metric mismatch and omitted prior-work comparison weaken support but do not make the derivation circular.

full rationale

The paper contains no formal derivation chain whose conclusion is equivalent to its premises. EchoSpot is trained end-to-end with transcription-only supervision and evaluated on standard external benchmarks (IC13, IC15, Total-Text, CTW1500); no parameter is fitted to a subset of data and then renamed as a prediction, and no uniqueness theorem or load-bearing self-citation is invoked to force the architecture. The recognition-attention activation map used as a localization signal is an intentionally bootstrapped pseudo-label loop, not a definitional identity: both recognition and localization are learned from data and are testable independently. The abstract's 'competitive performance' claim is weakened by the paper's own tables, because EchoSpot's headline numbers use the single-point metric while the fully-supervised baselines use the polygon metric, and the reported polygon-metric numbers are 10-20 points lower; however, this is a metric-mismatch/support deficiency, not a circularity. Likewise, the paper's novelty is undercut by the cited prior work Tang et al. (2022b), 'You can even annotate text with voice: Transcription-only-supervised text spotting', which already proposed transcription-only and voice-based annotation, but again this is a novelty/support concern rather than a circular step. The related-work survey in Sec. 2.2 also gestures at caption-based training without a citation, a missing-reference issue that does not affect circularity. No circular step can be quoted with a specific reduction, so the honest finding is no significant circularity.

Assumptions & free parameters 2 free parameters · 4 assumptions · 0 invented entities

The paper introduces no new theoretical entities. Its load-bearing assumptions are empirical: that transcription-only data can drive implicit localization, that attention maps are a valid proxy for position, that audio transcription is accurate enough, and that comparisons are apples-to-apples. These are standard, but they are not validated beyond the paper's thin experiments. The coarse-to-fine iteration count is an explicit hand-chosen free parameter.

free parameters (2)

Number of coarse-to-fine cross-attention iterations = 1
Chosen by hand. Table 1 shows 2 iterations give better end-to-end scores (66.3 vs 65.1 on Total-Text 'None'), yet all main experiments use 1, which suggests the choice is not principled and may be post hoc.
Attention activation map threshold for binary mask = Not reported
The coarse-to-fine mechanism requires binarizing the attention map, but no threshold or method is described. This is an unstated free parameter that affects localization quality.

assumptions (4)

domain assumption Transcription annotations alone provide enough signal to learn accurate text locations from image-text interaction.
This is the core premise of the method, introduced in Sec. 1 and 3.1. It is not proven and is directly contradicted in strength by the paper's own results, which are usually below methods with some location supervision.
domain assumption The recognition-stage attention activation map reliably localizes text instances without any geometric supervision.
Used in Sec. 3.2 and Figure 2 to produce the coarse mask and the refined position. No analysis is provided to show this map's precision, and the lower polygon-metric scores suggest the localization is imperfect.
domain assumption Speech-to-text conversion yields sufficiently accurate transcriptions for training.
Sec. 3.4 introduces audio annotation and Table 6 reports results from audio-derived labels, but the paper does not report the speech recognition accuracy or its effect on noisy labels.
domain assumption Training on SynthText and COCO-Text together with benchmark training sets follows the same protocol as the compared methods.
Sec. 4.1 states the training data but does not specify which subsets of COCO-Text are used or whether the baseline methods used identical data. This is necessary for a fair comparison.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Hear the Scene: Audio-Enhanced Text Spotting." pith.science (2026). https://pith.science/paper/DQ7WGH2J

@misc{pith2026241219504,
  author       = {Pith},
  title        = {Pith review of: Hear the Scene: Audio-Enhanced Text Spotting},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/DQ7WGH2J}},
  note         = {Machine review of arXiv:2412.19504}
}

read the original abstract

Recent advancements in scene text spotting have focused on end-to-end methodologies that heavily rely on precise location annotations, which are often costly and labor-intensive to procure. In this study, we introduce an innovative approach that leverages only transcription annotations for training text spotting models, substantially reducing the dependency on elaborate annotation processes. Our methodology employs a query-based paradigm that facilitates the learning of implicit location features through the interaction between text queries and image embeddings. These features are later refined during the text recognition phase using an attention activation map. Addressing the challenges associated with training a weakly-supervised model from scratch, we implement a circular curriculum learning strategy to enhance model convergence. Additionally, we introduce a coarse-to-fine cross-attention localization mechanism for more accurate text instance localization. Notably, our framework supports audio-based annotation, which significantly diminishes annotation time and provides an inclusive alternative for individuals with disabilities. Our approach achieves competitive performance against existing benchmarks, demonstrating that high accuracy in text spotting can be attained without extensive location annotations.

Figures

Figures reproduced from arXiv: 2412.19504 by the authors.

**Figure 2.** Overview of the proposed EchoSpot model architecture. The visual and con [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗

**Figure 3.** Qualitative results. Images are selected from ICDAR 2013 (first col.), SCUT [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

14 extracted references · 5 canonical work pages

[6]

SPTS: Single-Point Text Spotting

Dezhi Peng, Xinyu Wang, Yuliang Liu, Jiaxin Zhang, Mingxin Huang, Songxuan Lai, Shenggao Zhu, Jing Li, Dahua Lin, Chunhua Shen, et al. Spts: Single-point text spotting. arXiv preprint arXiv:2112.07917 ,

work page Pith review arXiv
[8]

Wenhao Sun, Xue-Mei Dong, Benlei Cui, and Jingqun Tang

URL https://arxiv.org/abs/2412.12974. Wenhao Sun, Xue-Mei Dong, Benlei Cui, and Jingqun Tang. Attentive eraser: Unleashing diffusion model’s object removal potential via self-attention redirection guidance. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 39, pages 20734– 20742,

arXiv
[9]

Optimal boxes: boosting end-to-end scene text recognition by adjusting annotated bounding boxes via reinforcement learning

Jingqun Tang, Wenming Qian, Luchuan Song, Xiena Dong, Lan Li, and Xiang Bai. Optimal boxes: boosting end-to-end scene text recognition by adjusting annotated bounding boxes via reinforcement learning. In European Conference on Computer Vision, pages 233–248. Springer, 2022a. Jingqun Tang, Su Qiao, Benlei Cui, Yuhang Ma, Sheng Zhang, and Dimitrios Kanoulas...

arXiv
[11]

URL https://ojs.aaai

doi: 10.1609/aaai.v34i07.6896. URL https://ojs.aaai. org/index.php/AAAI/article/view/6896. Jianhui Wang, Zhifei Yang, Yangfan He, Huixiong Zhang, Yuxuan Chen, and Jing- wei Huang. Mari: Material retrieval integration across domains. arXiv preprint arXiv:2503.08111, 2025b. Junqiao Wang, Zeng Zhang, Yangfan He, Yuyang Song, Tianyu Shi, Yuchen Li, Hengyuan X...

arXiv
[12]

Pgnet: Real-time arbitrarily-shaped text spotting with point gathering network

Pengfei Wang, Chengquan Zhang, Fei Qi, Shanshan Liu, Xiaoqiang Zhang, Pengyuan Lyu, Junyu Han, Jingtuo Liu, Errui Ding, and Guangming Shi. Pgnet: Real-time arbitrarily-shaped text spotting with point gathering network. AAAI. AAAI , pages 2782–2790, 2021a. Wenhai Wang, Enze Xie, Xiang Li, Xuebo Liu, Ding Liang, Yang Zhibo, Tong Lu, and Chunhua Shen. Pan++:...

arXiv
[13]

Detecting curve text in the wild: New dataset and new solution

Liu Yuliang, Jin Lianwen, Zhang Shuaitao, and Zhang Sheng. Detecting curve text in the wild: New dataset and new solution. arXiv preprint arXiv:1712.02170 ,

arXiv
[2013]

Icdar 2015 competition on robust reading

Dimosthenis Karatzas, Lluis Gomez-Bigorda, Anguelos Nicolaou, Suman Ghosh, Andrew Bagdanov, Masakazu Iwamura, Jiri Matas, Lukas Neumann, Vijay Ramaseshan Chan- drasekhar, Shijian Lu, et al. Icdar 2015 competition on robust reading. In ICDAR, pages 1156–1160,

work page 2015
[2015]

Llava-onevision: Easy visual task transfer

Bo Li, Yuanhan Zhang, Dong Guo, Renrui Zhang, Feng Li, Hao Zhang, Kaichen Zhang, Yanwei Li, Ziwei Liu, and Chunyuan Li. Llava-onevision: Easy visual task transfer. arXiv preprint arXiv:2408.03326 , 2024a. Bohao Li, Yuying Ge, Yi Chen, Yixiao Ge, Ruimao Zhang, and Ying Shan. Seed-bench- 2-plus: Benchmarking multimodal large language models with text-rich v...

arXiv 2020

Show all 14 references

[2017]

Tabpedia: Towards comprehensive visual table understanding with concept synergy

Weichao Zhao, Hao Feng, Qi Liu, Jingqun Tang, Binghong Wu, Lei Liao, Shu Wei, Yongjie Ye, Hao Liu, Wengang Zhou, et al. Tabpedia: Towards comprehensive visual table understanding with concept synergy. Advances in Neural Information Processing Systems, 37:7185–7212, 2024a. Zhen...

arXiv
[2018]

Y. Liu, H. Chen, C. Shen, T. He, and L. Wang. Abcnet: Real-time scene text spotting with adaptive bezier-curve network. In 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , 2020a. Yuliang Liu, Hao Chen, Chunhua Shen, Tong He, Lianwen Jin, and Liangwe...

2020
[2019]

Icdar 2013 robust reading competition

Dimosthenis Karatzas, Faisal Shafait, Seiichi Uchida, Masakazu Iwamura, Lluis Gomez i Bigorda, Sergi Robles Mestre, Joan Mas, David Fernandez Mota, Jon Almazan Al- mazan, and Lluis Pere De Las Heras. Icdar 2013 robust reading competition. In Proc. ICDAR, pages 1484–1493,

2013
[2020]

ISBN 978-3-030-58621-8

Springer International Publishing. ISBN 978-3-030-58621-8. X. Liu, L. Ding, Y. Shi, D. Chen, and J. Yan. Fots: Fast oriented text spotting with a unified network. In 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR),

2018
[2023]

Textsquare: Scaling up text-centric visual instruction tuning

Jingqun Tang, Chunhui Lin, Zhen Zhao, Shu Wei, Binghong Wu, Qi Liu, Hao Feng, Yang Li, Siqi Wang, Lei Liao, et al. Textsquare: Scaling up text-centric visual instruction tuning. arXiv preprint arXiv:2404.12803 , 2024a. Jingqun Tang, Qi Liu, Yongjie Ye, Jinghui Lu, Shu Wei, Chu...

arXiv
[2024]

URL https://arxiv.org/abs/2410.11538. B. Shi, B. Xiang, and Y. Cong. An end-to-end trainable neural network for image-based sequence recognition and its application to scene text recognition. IEEE Transactions on Pattern Analysis & Machine Intelligence , 39(11):2298–2304,

arXiv

Pith tools

Reviewed August 11, 2026 · model on record in the stance chip above.

[1] [6]

SPTS: Single-Point Text Spotting

Dezhi Peng, Xinyu Wang, Yuliang Liu, Jiaxin Zhang, Mingxin Huang, Songxuan Lai, Shenggao Zhu, Jing Li, Dahua Lin, Chunhua Shen, et al. Spts: Single-point text spotting. arXiv preprint arXiv:2112.07917 ,

work page Pith review arXiv

[2] [8]

Wenhao Sun, Xue-Mei Dong, Benlei Cui, and Jingqun Tang

URL https://arxiv.org/abs/2412.12974. Wenhao Sun, Xue-Mei Dong, Benlei Cui, and Jingqun Tang. Attentive eraser: Unleashing diffusion model’s object removal potential via self-attention redirection guidance. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 39, pages 20734– 20742,

arXiv

[3] [9]

Optimal boxes: boosting end-to-end scene text recognition by adjusting annotated bounding boxes via reinforcement learning

Jingqun Tang, Wenming Qian, Luchuan Song, Xiena Dong, Lan Li, and Xiang Bai. Optimal boxes: boosting end-to-end scene text recognition by adjusting annotated bounding boxes via reinforcement learning. In European Conference on Computer Vision, pages 233–248. Springer, 2022a. Jingqun Tang, Su Qiao, Benlei Cui, Yuhang Ma, Sheng Zhang, and Dimitrios Kanoulas...

arXiv

[4] [11]

URL https://ojs.aaai

doi: 10.1609/aaai.v34i07.6896. URL https://ojs.aaai. org/index.php/AAAI/article/view/6896. Jianhui Wang, Zhifei Yang, Yangfan He, Huixiong Zhang, Yuxuan Chen, and Jing- wei Huang. Mari: Material retrieval integration across domains. arXiv preprint arXiv:2503.08111, 2025b. Junqiao Wang, Zeng Zhang, Yangfan He, Yuyang Song, Tianyu Shi, Yuchen Li, Hengyuan X...

arXiv

[5] [12]

Pgnet: Real-time arbitrarily-shaped text spotting with point gathering network

Pengfei Wang, Chengquan Zhang, Fei Qi, Shanshan Liu, Xiaoqiang Zhang, Pengyuan Lyu, Junyu Han, Jingtuo Liu, Errui Ding, and Guangming Shi. Pgnet: Real-time arbitrarily-shaped text spotting with point gathering network. AAAI. AAAI , pages 2782–2790, 2021a. Wenhai Wang, Enze Xie, Xiang Li, Xuebo Liu, Ding Liang, Yang Zhibo, Tong Lu, and Chunhua Shen. Pan++:...

arXiv

[6] [13]

Detecting curve text in the wild: New dataset and new solution

Liu Yuliang, Jin Lianwen, Zhang Shuaitao, and Zhang Sheng. Detecting curve text in the wild: New dataset and new solution. arXiv preprint arXiv:1712.02170 ,

arXiv

[7] [2013]

Icdar 2015 competition on robust reading

Dimosthenis Karatzas, Lluis Gomez-Bigorda, Anguelos Nicolaou, Suman Ghosh, Andrew Bagdanov, Masakazu Iwamura, Jiri Matas, Lukas Neumann, Vijay Ramaseshan Chan- drasekhar, Shijian Lu, et al. Icdar 2015 competition on robust reading. In ICDAR, pages 1156–1160,

work page 2015

[8] [2015]

Llava-onevision: Easy visual task transfer

Bo Li, Yuanhan Zhang, Dong Guo, Renrui Zhang, Feng Li, Hao Zhang, Kaichen Zhang, Yanwei Li, Ziwei Liu, and Chunyuan Li. Llava-onevision: Easy visual task transfer. arXiv preprint arXiv:2408.03326 , 2024a. Bohao Li, Yuying Ge, Yi Chen, Yixiao Ge, Ruimao Zhang, and Ying Shan. Seed-bench- 2-plus: Benchmarking multimodal large language models with text-rich v...

arXiv 2020