Pith. sign in

REVIEW 4 major objections 6 minor 1 cited by

EventSTR: A Benchmark Dataset and Baselines for Event Stream based Scene Text Recognition

T0 review · 4 major / 6 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 The paper proposes event stream based scene text recognition as a new task, introduces the EventSTR benchmark of 9,928 high-definition event samples covering Chinese and English text, and presents SimC-ESTR, an LLM-based recognizer that…

desk verdict The EventSTR dataset is a genuinely new resource worth knowing about, but the paper's central claim that event cameras solve low-light and motion-blur text recognition is undercut by the dataset's own exclusion of those cases and by test-set tuning. read the letter →

arxiv 2502.09020 v1 pith:SP2YGKZV submitted 2025-02-13 cs.CV cs.AI

Xiao Wang , Jingtao Jiang , Dong Li , Futian Wang , Lin Zhu , Yaowei Wang , Yongyong Tian , Jin Tang This is my paper
classification cs.CVcs.AI
keywords eventcamerascenetextrecognitionbenchmarkdatasetlargelanguagemodelmemorymechanismglypherrorcorrectionChinesestream
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper introduces event stream based scene text recognition (STR) as a new task: reading text from the asynchronous brightness-change signals recorded by event cameras instead of from ordinary RGB images. To make the task testable, it contributes EventSTR, a benchmark of 9,928 high-definition (1280x720) event samples containing both Chinese and English text, together with fine-tuned baselines that future work can compare against. It also proposes SimC-ESTR, a large-language-model recognizer that converts event frames into visual tokens, augments them with a memory mechanism, and corrects errors using a database of visually similar glyphs. On EventSTR, SimC-ESTR reports BLEU-1 of 0.638, ahead of the best compared baseline at 0.584, and the paper argues this establishes the feasibility of event-based STR.

What carries the argument

The machinery has four pieces. Event frames are produced by stacking the asynchronous event stream into frames, following the convention of a prior high-resolution event tracking dataset, so existing STR architectures can be applied. SimC-ESTR then encodes each event frame with a large vision transformer, projects the features through a Q-former (a cross-attention module that distills visual features into a small set of learned query tokens aligned with the language model), and concatenates prompt, query, and visual tokens as input to a frozen 7B LLaMA-family language model. A memory module stores learnable pattern vectors, retrieves the top-K most similar patterns by cosine similarity for each input token, and adds their weighted average back to the features, which the paper shows helps when event frames are sparse or noisy. Finally, a glyph error correction module builds a database of visually similar characters and words, checks the LLM's first prediction character by character against that database, and re-prompts the LLM with candidate replacements, which corrects confusions such as Chinese characters that look alike.

What would settle it

Record a held-out test set of event clips in the excluded conditions—very low light and fast camera motion—annotate their text, and run SimC-ESTR; if its BLEU scores fall to the level of RGB baselines or below the reported 0.638, the benchmark's motivating advantage is not established.

Watch

Extended reading notes

Core claim

The paper's central claim is that scene text can be recognized from event camera streams, and that this deserves to be a benchmarked task of its own rather than an afterthought of RGB-based STR. To support the claim, the authors collected EventSTR, 9,928 high-definition (1280x720) event clips under varied lighting, motion, occlusion, scene type, and text orientation, with labels for both Chinese and English characters. They then show that a recognizer built for event input—SimC-ESTR, a frozen large language model fed with visual tokens from an event-frame encoder, memory-enhanced features, and a glyph-similarity correction pass—reaches BLEU-1 0.638 on the test split, against 0.584 for the strongest compared baseline and 0.450 for a well-known RGB STR model. The authors read this as evidence that event-based STR is feasible and that the dataset plus the model give future work a point of comparison.

Load-bearing premise

The load-bearing premise is that EventSTR's clips represent the low-light, motion-blur, and occlusion conditions that motivate event cameras, yet the annotation protocol explicitly excludes scenes that are excessively dark or motion-blurred.

Editorial extensions

If this is right

  • Event streams can serve as a viable input modality for scene text recognition, at least on the collected benchmark, giving an alternative to RGB frames.
  • Existing RGB-trained STR models transfer poorly to event data, so event-specific training data and models are needed rather than simple fine-tuning of frame-based recognizers.
  • The memory module's top-K retrieval improves recognition on noisy or incomplete event frames, suggesting memory-augmented visual features help when the input is sparse.
  • The glyph error correction module reduces confusion among visually similar Chinese characters and English words, and component analysis attributes most of the reported gain to it.
  • The benchmark's 6949/993/1986 split and released baselines provide a standard protocol for future event-based STR evaluation.

Reading between the lines

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

  • Because the annotation protocol discards excessively dark and motion-blurred scenes, the benchmark likely underrepresents the very conditions where event cameras should outperform RGB; the reported advantage may not transfer until such clips are included.
  • The two simulated evaluation sets (WordArt* and IC15*) are generated from RGB images with an event simulator, so they test recognition of simulated events rather than native event-camera recordings, and results on them should not be read as real-sensor performance.
  • The gains over RGB baselines may partly come from the frozen LLM's language prior and the manually curated glyph database rather than from event-specific visual information; an ablation feeding the same architecture RGB frames would separate these factors.
  • A natural extension the paper leaves implicit is using the raw asynchronous event stream or multi-frame stacks instead of a single stacked event frame, which could exploit temporal resolution more fully.
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 / 6 minor

Summary. The paper introduces EventSTR, a dataset of 9,928 event-camera recordings of scene text (1280x720, Chinese and English), and proposes SimC-ESTR, an LLM-based recognition framework that augments vision tokens with a memory module and applies a glyph-based error correction step. The authors benchmark several existing STR methods on EventSTR and on two simulated event datasets (WordArt* and IC15*), reporting that SimC-ESTR achieves the best BLEU scores on EventSTR (BLEU-1 0.638) but lower accuracy than several baselines on the simulated datasets. The central claims are that EventSTR is the first event-based scene text recognition benchmark and that SimC-ESTR is an effective baseline for it.

Significance. If the validity concerns identified below were resolved, EventSTR could be a useful first benchmark for event-based scene text recognition, and the paper provides a substantial data collection effort plus multiple baseline comparisons. The proposed framework is also a reasonable LLM-based starting point, and the paper honestly lists several limitations in Section V-G. However, the significance is currently undercut by an internal contradiction in the data collection protocol, the discarding of temporal information in the event representation, and the selection of key components on the test set. These issues prevent the results from supporting the paper's motivating claim that event cameras offer an advantage for text recognition under low-light and motion-blur conditions.

major comments (4)
  1. [Section IV-B] The annotation protocol states 'No annotations are made for scenes that are excessively dark or have motion blur, as these conditions hinder text recognition.' This directly contradicts the dataset's stated collection protocols in Section IV-A (low-light capture and motion variability) and the motivation in the Introduction and Fig. 1, which argue that event cameras excel precisely in these conditions. As a result, the test set used for Table II does not contain the challenging scenarios that justify event-based STR, and the reported BLEU scores cannot be taken as evidence that event cameras improve text recognition under low illumination or motion blur.
  2. [Section IV-C] The dataset representation discards temporal information: Section IV-C states that each video sequence is converted into 19 event frames and 'the first frame selected as the final representation of the dataset.' All models are therefore trained and evaluated on a single event frame, not on an event stream. Since high temporal resolution is the defining advantage of event cameras, this choice undermines the paper's claim to introduce 'event stream based' scene text recognition. The authors should either evaluate multiple frame counts or temporal aggregation strategies, or reframe the contribution as event-frame-based STR.
  3. [Section III-C and Section V-E] The Glyph Error Correction Module is 'task-specifically adjusted' based on recognition errors observed in preliminary experiments on the same EventSTR data, and the hyperparameters—top-K=64 in the memory module, 10 similar-word candidates, and Prompt 3—are each selected by comparing BLEU scores on the test split (Tables V, VI, VII). This constitutes test-set tuning and likely inflates the reported improvements. The 0.638 BLEU-1 result is therefore not an unbiased estimate of the method's generalization. All ablations should be performed on the validation split, and the glyph database construction should be frozen before any test-set evaluation.
  4. [Section V-C and Table III] On the WordArt* and IC15* simulated datasets, SimC-ESTR underperforms several standard STR models (e.g., 65.1 vs 75.0 for PARSeq on WordArt*, and 56.8 vs 72.7 on IC15*). The acknowledgment that VQA pre-training is not optimal for OCR does not explain why the proposed memory and correction modules fail to improve these results. Since the method's effectiveness claim rests almost entirely on the EventSTR results, which are compromised by the test-set tuning described above, the paper's central experimental evidence is not convincing.
minor comments (6)
  1. [Section V-C and Table II] The text reports SimC-ESTR BLEU scores as 0.629, 0.570, 0.486, and 0.417, but Table II lists 0.638, 0.583, 0.500, and 0.430. The text numbers correspond to the GECM-only configuration in Table IV or Prompt 3 in Table VI; please correct this inconsistency.
  2. [Section IV-C] The dataset is described both as 9,928 'video sequences' and as 9,928 'event samples'; please clarify whether the 6,949/993/1,986 train/val/test split refers to the original sequences or to the first-frame representations.
  3. [Section V-B] It is unclear whether SimC-ESTR is fine-tuned separately on each dataset or only on EventSTR; please specify the fine-tuning protocol for WordArt* and IC15* so that comparisons in Table III are fair.
  4. [Section V-A] Please explain how BLEU is computed for EventSTR scenes that contain multiple text lines or words, including the tokenization procedure for mixed Chinese/English text and how multiple ground-truth strings are handled.
  5. [Abstract] The abstract promises release of source code and pre-trained models but does not mention releasing the EventSTR dataset itself; for a benchmark paper, a clear data-availability statement is essential.
  6. [Section III-C] The 'Glyph Error Correction Module' is a database-based post-processing step combined with a prompt update, not a learned module; consider renaming it to 'Glyph Error Correction Mechanism' or similar to avoid confusion.

Circularity Check

2 steps flagged · score 6.0 of 10

SimC-ESTR's glyph-correction database and key hyperparameters are fitted to the EventSTR evaluation benchmark, so the headline EventSTR gains are partially circular; the dataset itself and the WordArt*/IC15* results remain independent.

  1. fitted input called prediction [Section III-C, 'Glyph Error Correction Module', '1) Similar Glyph Database Construction'; evaluated in Section V-D Table IV and Section V-F visualizations.]
    "Task-Specific Adjustment: Modifying the database according to the recognition errors observed in preliminary experiments. This helps optimize the database for event-based scene text recognition scenarios, enhancing the correction module’s performance."

    The similar-glyph database is explicitly modified using recognition errors observed in preliminary experiments, and the corrected outputs are then reported as final recognition results on the EventSTR benchmark (Table IV attributes a +0.045 BLEU-1 gain to GECM; Figure 7 showcases corrections of exactly this kind, e.g., '才' corrected to '力' and '里' to '偶'). The component analysis and visualizations use EventSTR, and the final BLEU values in Table IV match the test-set values in Table II, so the database is calibrated on the same benchmark whose test performance is then presented as the model's predictive result.

  2. fitted input called prediction [Section V-E, 'Impact of Top-K Selection in Memory Module', 'Analysis of Different Prompts for Error Correction', and 'Analysis of the Size of the Similar Word Database' (Tables V-VII).]
    "From the results, we observe that increasing the number of candidate words from 5 to 7 provides a slight improvement in BLEU scores across all metrics. However, the most significant gain in performance is achieved when the database size is increased to 10 candidates per glyph, where the BLEU-1, BLEU-2, BLEU-3, and BLEU-4 scores reach their highest values."

    The paper selects the prompt phrasing, the memory top-K value, and the number of candidate glyphs by comparing BLEU scores, then reports the selected configuration as the final SimC-ESTR result. The chosen settings (Prompt 3, K=64, 10 candidates) produce BLEU-1 values of 0.629, 0.638, and 0.629 respectively, and the final model with both modules reaches 0.638 in Table II and Table IV. Since these ablations are run on the EventSTR benchmark that is also the test set for the headline SOTA comparison, the hyperparameters are fit to the evaluation data. The claimed advantage over BLIVA is therefore partly a consequence of test-set selection rather than a validated, transferable design choice.

full rationale

The circularity is real but confined to the EventSTR headline evaluation. The dataset contribution itself is not circular: EventSTR is a newly collected artifact with documented statistics, splits, and external baselines, and those baselines do not reduce to the paper's claims. The WordArt* and IC15* simulations are independent external benchmarks, and on those the proposed method underperforms PARSeq, which is consistent with a non-circular (if less favorable) experimental outcome. However, the central EventSTR result is partially circular in two connected ways. First, the Glyph Error Correction Module's similar-glyph database is explicitly 'task-specifically adjusted' based on recognition errors observed in preliminary experiments, and the same EventSTR benchmark is then used to demonstrate the module's improvement. Second, the prompt wording, the number of candidate glyphs, and the memory top-K are all chosen by comparing BLEU scores on EventSTR, with the selected values coinciding with the final reported SOTA numbers. Thus the +0.054 BLEU-1 gain credited to GECM and MM in Table IV is not a clean out-of-sample prediction; it is partially fitted to the evaluation set. Separately, the Section IV-B annotation protocol excludes excessively dark and motion-blurred scenes ('No annotations are made for scenes that are excessively dark or have motion blur'), which undermines the motivational claim that event cameras help in exactly those regimes, but that is a dataset-validity and consistency concern rather than a circular-derivation issue, so it is not counted in the score. Overall, the derivation chain is not fully circular because the visual encoder/LLM training and the external WordArt*/IC15* evaluations provide independent content, but the headline EventSTR 'prediction' is materially fitted, warranting a score of 6.

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

The central claims rest on several untested assumptions about event-frame representations and domain transfer, plus a set of hyperparameters and a glyph database that are effectively fitted to the test set. The method contributes no new physical or conceptual entities beyond standard learned modules.

free parameters (5)
  • Top-K in memory module = 64
    Selected by ablation on EventSTR test BLEU scores (Table V); K=64 maximizes BLEU-1.
  • Number of similar word candidates per glyph = 10
    Selected by ablation on the test set (Table VII); 10 candidates maximizes BLEU-1.
  • Glyph correction prompt phrasing = Prompt 3
    Selected by ablation on test-set BLEU (Table VI); Prompt 3 outperforms other formulations.
  • Similar glyph database contents = Hand-built list of visually similar Chinese and English characters
    The database is 'task-specifically adjusted' based on errors observed in preliminary experiments, meaning it is fitted to the dataset distribution.
  • Event frame count and selection = 19 frames, first frame used
    Each event sequence is converted into 19 frames and the first frame is selected as the sample (Section IV-C), a representation choice without validation.
assumptions (5)
  • domain assumption Event frames obtained by stacking events preserve enough text structure for recognition.
    The whole pipeline operates on stacked event frames rather than raw event streams; the paper does not validate that this representation retains text legibility (Section III-B).
  • ad hoc to paper The first frame of a 19-frame stack is a sufficient representation of each event sequence for text recognition.
    Section IV-C states the first frame is selected as the final representation without comparing to other frames or aggregating them.
  • ad hoc to paper Excluding excessively dark and motion-blurred scenes from annotation does not bias the benchmark away from the claimed advantage.
    Section IV-B excludes such scenes for annotation, contradicting the paper's motivation that event cameras handle low light and motion blur.
  • domain assumption The Vicuna LLM can correct OCR errors using candidate lists derived from a hand-built glyph database.
    The Glyph Error Correction Module effectiveness rests on the LLM following the correction prompt (Section III-C.2), and this is never validated without the candidate lists.
  • domain assumption EVA-CLIP visual features pretrained on RGB images transfer to event-frame images.
    The visual encoder is initialized from EVA-CLIP; no domain adaptation or validation on event data is shown before fine-tuning (Section V-B).

how reviews work

0 comments
Cite this review

Pith. "Pith review of EventSTR: A Benchmark Dataset and Baselines for Event Stream based Scene Text Recognition." pith.science (2026). https://pith.science/paper/SP2YGKZV

@misc{pith2026250209020,
  author       = {Pith},
  title        = {Pith review of: EventSTR: A Benchmark Dataset and Baselines for Event Stream based Scene Text Recognition},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/SP2YGKZV}},
  note         = {Machine review of arXiv:2502.09020}
}
read the original abstract

Mainstream Scene Text Recognition (STR) algorithms are developed based on RGB cameras which are sensitive to challenging factors such as low illumination, motion blur, and cluttered backgrounds. In this paper, we propose to recognize the scene text using bio-inspired event cameras by collecting and annotating a large-scale benchmark dataset, termed EventSTR. It contains 9,928 high-definition (1280 * 720) event samples and involves both Chinese and English characters. We also benchmark multiple STR algorithms as the baselines for future works to compare. In addition, we propose a new event-based scene text recognition framework, termed SimC-ESTR. It first extracts the event features using a visual encoder and projects them into tokens using a Q-former module. More importantly, we propose to augment the vision tokens based on a memory mechanism before feeding into the large language models. A similarity-based error correction mechanism is embedded within the large language model to correct potential minor errors fundamentally based on contextual information. Extensive experiments on the newly proposed EventSTR dataset and two simulation STR datasets fully demonstrate the effectiveness of our proposed model. We believe that the dataset and algorithmic model can innovatively propose an event-based STR task and are expected to accelerate the application of event cameras in various industries. The source code and pre-trained models will be released on https://github.com/Event-AHU/EventSTR

Figures

Figures reproduced from arXiv: 2502.09020 by the authors.

Figure 1
Figure 1. Examples illustrating the motivation behind EventSTR. (a) Challenges of scene text recognition under low-light conditions where RGB cameras struggle [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. An overview of our proposed large language model based event stream scene text recognition framework, termed SimC-ESTR. Given the event streams, [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Illustration of some representative samples of our proposed EventSTR dataset. The left side displays the event stream, while the right side shows the [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figures from the paper (4 more)
Figure 4
Figure 4. Figure 4: Statistical analysis for the EventSTR dataset. (a) The number of images with different text lengths. (b) Distribution of the number of characters. [PITH_FULL_IMAGE:figures/full_fig_p006_4.png]
Figure 5
Figure 5. Figure 5: The word cloud visually represents the frequency distribution of [PITH_FULL_IMAGE:figures/full_fig_p007_5.png]
Figure 6
Figure 6. Figure 6: Illustration of representative samples of the synthetic WordArt* and IC15* dataset. [PITH_FULL_IMAGE:figures/full_fig_p008_6.png]
Figure 7
Figure 7. Figure 7: Comparison of Baseline and Glyph-Corrected Recognition Results. [PITH_FULL_IMAGE:figures/full_fig_p011_7.png]

Discussion (0). Sign in to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. ESTR-CoT: Towards Explainable and Accurate Event Stream based Scene Text Recognition with Chain-of-Thought Reasoning

    cs.CV 2025-07 conditional novelty 5.0 of 10

    An event-stream scene text recognizer trained with LLM-generated chain-of-thought rationales improves BLEU-1 on EventSTR from 0.638 to 0.648 and accuracy on WordArt* and IC15* by about half a point.

Reference graph

Works this paper leans on

57 extracted references · 17 canonical work pages · cited by 1 Pith paper

  1. [1]

    Scene text recognition with permuted autoregressive sequence models,

    D. Bautista and R. Atienza, “Scene text recognition with permuted autoregressive sequence models,” in European conference on computer vision. Springer, 2022, pp. 178–196

    2022

  2. [2]

    Multi-granularity prediction for scene text recognition,

    P. Wang, C. Da, and C. Yao, “Multi-granularity prediction for scene text recognition,” in European Conference on Computer Vision . Springer, 2022, pp. 339–355

    2022

  3. [3]

    mplug-docowl 1.5: Unified structure learning for ocr-free document understanding,

    A. Hu, H. Xu, J. Ye, M. Yan, L. Zhang, B. Zhang, C. Li, J. Zhang, Q. Jin, F. Huang et al. , “mplug-docowl 1.5: Unified structure learning for ocr-free document understanding,” arXiv preprint arXiv:2403.12895, 2024

    arXiv 2024

  4. [4]

    Textmonkey: An ocr-free large multimodal model for understanding document,

    Y . Liu, B. Yang, Q. Liu, Z. Li, Z. Ma, S. Zhang, and X. Bai, “Textmonkey: An ocr-free large multimodal model for understanding document,” arXiv preprint arXiv:2403.04473 , 2024

    arXiv 2024

  5. [5]

    Docpedia: Un- leashing the power of large multimodal model in the frequency domain for versatile document understanding,

    H. Feng, Q. Liu, H. Liu, W. Zhou, H. Li, and C. Huang, “Docpedia: Un- leashing the power of large multimodal model in the frequency domain for versatile document understanding,”arXiv preprint arXiv:2311.11810, 2023

    arXiv 2023

  6. [6]

    Asynchronous spatio-temporal memory network for continuous event-based object detection,

    J. Li, J. Li, L. Zhu, X. Xiang, T. Huang, and Y . Tian, “Asynchronous spatio-temporal memory network for continuous event-based object detection,” IEEE Transactions on Image Processing , vol. 31, pp. 2975– 2987, 2022

    work page 2022

  7. [7]

    Event stream-based visual object tracking: A high-resolution bench- mark dataset and a novel baseline,

    X. Wang, S. Wang, C. Tang, L. Zhu, B. Jiang, Y . Tian, and J. Tang, “Event stream-based visual object tracking: A high-resolution bench- mark dataset and a novel baseline,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , 2024, pp. 19 248–19 257

    2024

  8. [8]

    MambaEVT: Event Stream based Visual Object Tracking using State Space Model

    X. Wang, S. Wang, X. Wang, Z. Zhao, L. Zhu, B. Jiang et al. , “Mambaevt: Event stream based visual object tracking using state space model,” arXiv preprint arXiv:2408.10487 , 2024

    work page Pith review arXiv 2024

Show all 57 references
  1. [9]

    Semantic-aware frame-event fusion based pattern recognition via large vision–language models,

    D. Li, J. Jin, Y . Zhang, Y . Zhong, Y . Wu, L. Chen, X. Wang, and B. Luo, “Semantic-aware frame-event fusion based pattern recognition via large vision–language models,” Pattern Recognition , vol. 158, p. 111080, 2025

    2025

  2. [10]

    Hardvs: Revisiting human activity recognition with dynamic vision sensors,

    X. Wang, Z. Wu, B. Jiang, Z. Bao, L. Zhu, G. Li, Y . Wang, and Y . Tian, “Hardvs: Revisiting human activity recognition with dynamic vision sensors,” in Proceedings of the AAAI Conference on Artificial Intelligence, vol. 38, no. 6, 2024, pp. 5615–5623

    2024

  3. [11]

    Event-based semantic segmentation with posterior attention,

    Z. Jia, K. You, W. He, Y . Tian, Y . Feng, Y . Wang, X. Jia, Y . Lou, J. Zhang, G. Li et al., “Event-based semantic segmentation with posterior attention,” IEEE Transactions on Image Processing , vol. 32, pp. 1829– 1842, 2023

    2023

  4. [12]

    Scene text detection and recognition: The deep learning era,

    S. Long, X. He, and C. Yao, “Scene text detection and recognition: The deep learning era,” International Journal of Computer Vision , vol. 129, no. 1, pp. 161–184, 2021

    2021

  5. [13]

    Large-scale multi-modal pre-trained models: A comprehensive survey,

    X. Wang, G. Chen, G. Qian, P. Gao, X.-Y . Wei, Y . Wang, Y . Tian, and W. Gao, “Large-scale multi-modal pre-trained models: A comprehensive survey,” Machine Intelligence Research , vol. 20, no. 4, pp. 447–482, 2023

    2023

  6. [16]

    An end-to-end trainable neural network for image-based sequence recognition and its application to scene text recognition,

    B. Shi, X. Bai, and C. Yao, “An end-to-end trainable neural network for image-based sequence recognition and its application to scene text recognition,” IEEE transactions on pattern analysis and machine intelligence, vol. 39, no. 11, pp. 2298–2304, 2016

    2016

  7. [17]

    Scene text recognition from two-dimensional perspective,

    M. Liao, J. Zhang, Z. Wan, F. Xie, J. Liang, P. Lyu, C. Yao, and X. Bai, “Scene text recognition from two-dimensional perspective,” in Proceedings of the AAAI conference on artificial intelligence , vol. 33, no. 01, 2019, pp. 8714–8721

    2019

  8. [18]

    Spotlight text detector: Spotlight on candidate regions like a camera,

    X. Han, J. Gao, C. Yang, Y . Yuan, and Q. Wang, “Spotlight text detector: Spotlight on candidate regions like a camera,” IEEE Transactions on Multimedia, 2024

    2024

  9. [19]

    Multi-modal in-context learning makes an ego-evolving scene text recognizer,

    Z. Zhao, J. Tang, C. Lin, B. Wu, C. Huang, H. Liu, X. Tan, Z. Zhang, and Y . Xie, “Multi-modal in-context learning makes an ego-evolving scene text recognizer,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , 2024, pp. 15 567–15 576

    2024

  10. [20]

    Self- supervised character-to-character distillation for text recognition,

    T. Guan, W. Shen, X. Yang, Q. Feng, Z. Jiang, and X. Yang, “Self- supervised character-to-character distillation for text recognition,” in 2023 IEEE/CVF International Conference on Computer Vision (ICCV) , 2023, pp. 19 416–19 427

    2023

  11. [21]

    Self- supervised implicit glyph attention for text recognition,

    T. Guan, C. Gu, J. Tu, X. Yang, Q. Feng, Y . Zhao, and W. Shen, “Self- supervised implicit glyph attention for text recognition,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recogni- tion, 2023, pp. 15 285–15 294

    2023

  12. [22]

    Cdistnet: Perceiving multi-domain character distance for robust text recognition,

    T. Zheng, Z. Chen, S. Fang, H. Xie, and Y .-G. Jiang, “Cdistnet: Perceiving multi-domain character distance for robust text recognition,” International Journal of Computer Vision , vol. 132, no. 2, pp. 300–318, 2024

    2024

  13. [23]

    V olter: Visual collaboration and dual-stream fusion for scene text recognition,

    J.-N. Li, X.-Q. Liu, X. Luo, and X.-S. Xu, “V olter: Visual collaboration and dual-stream fusion for scene text recognition,” IEEE Transactions on Multimedia, 2024

    2024

  14. [24]

    Image as a language: Revisiting scene text recognition via IEEE TRANSACTIONS ON ***, 2025 12 balanced, unified and synchronized vision-language reasoning network,

    J. Wei, H. Zhan, Y . Lu, X. Tu, B. Yin, C. Liu, and U. Pal, “Image as a language: Revisiting scene text recognition via IEEE TRANSACTIONS ON ***, 2025 12 balanced, unified and synchronized vision-language reasoning network,” Proceedings of the AAAI Conference on Artificial Int...

    2025

  15. [25]

    Multi-modal text recognition networks: Interactive enhancements between visual and semantic features,

    B. Na, Y . Kim, and S. Park, “Multi-modal text recognition networks: Interactive enhancements between visual and semantic features,” in European Conference on Computer Vision . Springer, 2022, pp. 446– 463

    2022

  16. [26]

    Levenshtein ocr,

    C. Da, P. Wang, and C. Yao, “Levenshtein ocr,” in European Conference on Computer Vision . Springer, 2022, pp. 322–338

    2022

  17. [27]

    Read like humans: Autonomous, bidirectional and iterative language modeling for scene text recognition,

    S. Fang, H. Xie, Y . Wang, Z. Mao, and Y . Zhang, “Read like humans: Autonomous, bidirectional and iterative language modeling for scene text recognition,” in Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , 2021, pp. 7098–7107

    2021

  18. [28]

    A survey of large language models,

    W. X. Zhao, K. Zhou, J. Li, T. Tang, X. Wang, Y . Hou, Y . Min, B. Zhang, J. Zhang, Z. Dong et al. , “A survey of large language models,” arXiv preprint arXiv:2303.18223, 2023

    2023 arXiv

  19. [29]

    Vary: Scaling up the vision vocabulary for large vision-language model,

    H. Wei, L. Kong, J. Chen, L. Zhao, Z. Ge, J. Yang, J. Sun, C. Han, and X. Zhang, “Vary: Scaling up the vision vocabulary for large vision-language model,” in European Conference on Computer Vision . Springer, 2025, pp. 408–424

    2025

  20. [30]

    General ocr theory: Towards ocr-2.0 via a unified end-to-end model,

    H. Wei, C. Liu, J. Chen, J. Wang, L. Kong, Y . Xu, Z. Ge, L. Zhao, J. Sun, Y . Penget al., “General ocr theory: Towards ocr-2.0 via a unified end-to-end model,” arXiv preprint arXiv:2409.01704 , 2024

    2024 arXiv

  21. [31]

    Event- based vision: A survey,

    G. Gallego, T. Delbrück, G. Orchard, C. Bartolozzi, B. Taba, A. Censi, S. Leutenegger, A. J. Davison, J. Conradt, K. Daniilidis et al., “Event- based vision: A survey,” IEEE transactions on pattern analysis and machine intelligence, vol. 44, no. 1, pp. 154–180, 2020

    2020

  22. [32]

    Evcslr: Event-guided continuous sign language recognition and benchmark,

    Y . Jiang, Y . Wang, S. Li, Y . Zhang, Q. Guo, Q. Chu, and Y . Gao, “Evcslr: Event-guided continuous sign language recognition and benchmark,” IEEE Transactions on Multimedia , 2024

    2024

  23. [33]

    Recurrent vision transformers for object detection with event cameras,

    M. Gehrig and D. Scaramuzza, “Recurrent vision transformers for object detection with event cameras,” in Proceedings of the IEEE/CVF confer- ence on computer vision and pattern recognition , 2023, pp. 13 884– 13 893

    2023

  24. [34]

    Masked autoen- coders in 3d point cloud representation learning,

    J. Jiang, X. Lu, L. Zhao, R. Dazaley, and M. Wang, “Masked autoen- coders in 3d point cloud representation learning,” IEEE Transactions on Multimedia, 2023

    2023

  25. [35]

    Llama: Open and efficient foundation language models,

    H. Touvron, T. Lavril, G. Izacard, X. Martinet, M.-A. Lachaux, T. Lacroix, B. Rozière, N. Goyal, E. Hambro, F. Azhar et al. , “Llama: Open and efficient foundation language models,” arXiv preprint arXiv:2302.13971, 2023

    2023 arXiv

  26. [36]

    Eva-clip: Improved training techniques for clip at scale,

    Q. Sun, Y . Fang, L. Wu, X. Wang, and Y . Cao, “Eva-clip: Improved training techniques for clip at scale,” arXiv preprint arXiv:2303.15389 , 2023

    2023 arXiv

  27. [37]

    Synthetic data and artificial neural networks for natural scene text recognition,

    M. Jaderberg, K. Simonyan, A. Vedaldi, and A. Zisserman, “Synthetic data and artificial neural networks for natural scene text recognition,” arXiv preprint arXiv:1406.2227 , 2014

    2014 arXiv

  28. [38]

    Synthetic data for text localisation in natural images,

    A. Gupta, A. Vedaldi, and A. Zisserman, “Synthetic data for text localisation in natural images,” in Proceedings of the IEEE conference on computer vision and pattern recognition , 2016, pp. 2315–2324

    2016

  29. [39]

    End-to-end scene text recog- nition,

    K. Wang, B. Babenko, and S. Belongie, “End-to-end scene text recog- nition,” in 2011 International conference on computer vision . IEEE, 2011, pp. 1457–1464

    2011

  30. [40]

    Scene text recognition using higher order language priors,

    A. Mishra, K. Alahari, and C. Jawahar, “Scene text recognition using higher order language priors,” in BMVC-British machine vision confer- ence. BMV A, 2012

    2012

  31. [41]

    Icdar 2013 robust reading competition,

    D. Karatzas, F. Shafait, S. Uchida, M. Iwamura, L. G. i Bigorda, S. R. Mestre, J. Mas, D. F. Mota, J. A. Almazan, and L. P. De Las Heras, “Icdar 2013 robust reading competition,” in 2013 12th international conference on document analysis and recognition . IEEE, 2013, pp. 1484–1493

    2013

  32. [42]

    Recognizing text with perspective distortion in natural scenes,

    T. Q. Phan, P. Shivakumara, S. Tian, and C. L. Tan, “Recognizing text with perspective distortion in natural scenes,” inProceedings of the IEEE international conference on computer vision , 2013, pp. 569–576

    2013

  33. [43]

    A robust arbitrary text detection system for natural scene images,

    A. Risnumawan, P. Shivakumara, C. S. Chan, and C. L. Tan, “A robust arbitrary text detection system for natural scene images,” Expert Systems with Applications, vol. 41, no. 18, pp. 8027–8048, 2014

    2014

  34. [44]

    Icdar 2015 competition on robust reading,

    D. Karatzas, L. Gomez-Bigorda, A. Nicolaou, S. Ghosh, A. Bagdanov, M. Iwamura, J. Matas, L. Neumann, V . R. Chandrasekhar, S. Lu et al., “Icdar 2015 competition on robust reading,” in 2015 13th international conference on document analysis and recognition (ICDAR) . IEEE, 2015,...

    2015

  35. [45]

    Coco-text: Dataset and benchmark for text detection and recognition in natural images,

    A. Veit, T. Matera, L. Neumann, J. Matas, and S. Belongie, “Coco-text: Dataset and benchmark for text detection and recognition in natural images,” arXiv preprint arXiv:1601.07140 , 2016

    2016 arXiv

  36. [46]

    Icdar2017 competition on reading chinese text in the wild (rctw- 17),

    B. Shi, C. Yao, M. Liao, M. Yang, P. Xu, L. Cui, S. Belongie, S. Lu, and X. Bai, “Icdar2017 competition on reading chinese text in the wild (rctw- 17),” in 2017 14th iapr international conference on document analysis and recognition (ICDAR), vol. 1. IEEE, 2017, pp. 1429–1434

    2017

  37. [47]

    Uber-text: A large-scale dataset for optical character recognition from street-level imagery,

    Y . Zhang, L. Gueguen, I. Zharkov, P. Zhang, K. Seifert, and B. Kadlec, “Uber-text: A large-scale dataset for optical character recognition from street-level imagery,” in SUNw: Scene Understanding Workshop-CVPR, vol. 2017, 2017, p. 5

    2017

  38. [48]

    Icdar2019 robust reading challenge on arbitrary- shaped text-rrc-art,

    C. K. Chng, Y . Liu, Y . Sun, C. C. Ng, C. Luo, Z. Ni, C. Fang, S. Zhang, J. Han, E. Ding et al., “Icdar2019 robust reading challenge on arbitrary- shaped text-rrc-art,” in 2019 International Conference on Document Analysis and Recognition (ICDAR) . IEEE, 2019, pp. 1571–1576

    2019

  39. [49]

    Icdar 2019 robust reading challenge on reading chinese text on signboard,

    R. Zhang, Y . Zhou, Q. Jiang, Q. Song, N. Li, K. Zhou, L. Wang, D. Wang, M. Liao, M. Yang et al., “Icdar 2019 robust reading challenge on reading chinese text on signboard,” in 2019 international conference on document analysis and recognition (ICDAR). IEEE, 2019, pp. 1577– 1581

    2019

  40. [50]

    Icdar 2019 competition on large-scale street view text with partial labeling-rrc-lsvt,

    Y . Sun, Z. Ni, C.-K. Chng, Y . Liu, C. Luo, C. C. Ng, J. Han, E. Ding, J. Liu, D. Karatzas et al. , “Icdar 2019 competition on large-scale street view text with partial labeling-rrc-lsvt,” in 2019 International Conference on Document Analysis and Recognition (ICDAR) . IEEE, 2...

    2019

  41. [51]

    Icdar2019 robust reading challenge on multi-lingual scene text detection and recognition—rrc- mlt-2019,

    N. Nayef, Y . Patel, M. Busta, P. N. Chowdhury, D. Karatzas, W. Khlif, J. Matas, U. Pal, J.-C. Burie, C.-l. Liu et al., “Icdar2019 robust reading challenge on multi-lingual scene text detection and recognition—rrc- mlt-2019,” in 2019 International conference on document analys...

    2019

  42. [52]

    Textocr: Towards large-scale end-to-end reasoning for arbitrary-shaped scene text,

    A. Singh, G. Pang, M. Toh, J. Huang, W. Galuba, and T. Hassner, “Textocr: Towards large-scale end-to-end reasoning for arbitrary-shaped scene text,” in Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , 2021, pp. 8802–8812

    2021

  43. [53]

    Toward understand- ing wordart: Corner-guided transformer for scene text recognition,

    X. Xie, L. Fu, Z. Zhang, Z. Wang, and X. Bai, “Toward understand- ing wordart: Corner-guided transformer for scene text recognition,” in European conference on computer vision. Springer, 2022, pp. 303–321

    2022

  44. [54]

    Revisiting scene text recognition: A data perspective,

    Q. Jiang, J. Wang, D. Peng, C. Liu, and L. Jin, “Revisiting scene text recognition: A data perspective,” in Proceedings of the IEEE/CVF international conference on computer vision , 2023, pp. 20 543–20 554

    2023

  45. [55]

    Lister: Neighbor decoding for length-insensitive scene text recognition,

    C. Cheng, P. Wang, C. Da, Q. Zheng, and C. Yao, “Lister: Neighbor decoding for length-insensitive scene text recognition,” in Proceedings of the IEEE/CVF International Conference on Computer Vision , 2023, pp. 19 541–19 551

    2023

  46. [56]

    Reading and writing: Discriminative and generative modeling for self-supervised text recognition,

    M. Yang, M. Liao, P. Lu, J. Wang, S. Zhu, H. Luo, Q. Tian, and X. Bai, “Reading and writing: Discriminative and generative modeling for self-supervised text recognition,” in Proceedings of the 30th ACM International Conference on Multimedia , 2022, pp. 4214–4223

    2022

  47. [57]

    Esim: an open event camera simulator,

    H. Rebecq, D. Gehrig, and D. Scaramuzza, “Esim: an open event camera simulator,” in Conference on robot learning . PMLR, 2018, pp. 969– 982

    2018

  48. [58]

    Bliva: A simple multimodal llm for better handling of text-rich visual questions,

    W. Hu, Y . Xu, Y . Li, W. Li, Z. Chen, and Z. Tu, “Bliva: A simple multimodal llm for better handling of text-rich visual questions,” in Proceedings of the AAAI Conference on Artificial Intelligence , vol. 38, no. 3, 2024, pp. 2256–2264

    2024

  49. [59]

    Decoupled weight decay regularization,

    I. Loshchilov, “Decoupled weight decay regularization,” arXiv preprint arXiv:1711.05101, 2017

    2017 arXiv

Pith tools

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