Pith. sign in

REVIEW 5 major objections 4 minor 1 cited by

The Devil is in Fine-tuning and Long-tailed Problems:A New Benchmark for Scene Text Detection

T0 review · 5 major / 4 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 Scene-text detector scores on academic benchmarks are inflated by per-dataset fine-tuning and by a long tail of rare text types that benchmarks ignore.

desk verdict A genuinely useful long-tailed scene text benchmark, but the unreported IoU threshold in its construction and the overstated causal claim need fixing before I'd trust the central narrative. read the letter →

arxiv 2505.15649 v1 pith:AQSBO7BX submitted 2025-05-21 cs.CV

Tianjiao Cao , Jiahao Lyu , Weichao Zeng , Weimin Mu , Yu Zhou This is my paper · ORCID
classification cs.CV
keywords scenetextdetectionlong-taileddistributionbenchmarkconstructiondataset-specificoptimizationjoint-datasetlearningself-supervisedmaskedautoencoderfine-tuninggap
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 argues that the gap between scene-text detectors' near-saturation scores on academic benchmarks and their uneven behavior in practice has two analyzable sources. The first is the Fine-tuning Gap: under Dataset-Specific Optimization (DSO), a model is fine-tuned on one dataset's training split and tested on that same dataset's test split, so benchmark scores reward overfitting to dataset-specific biases. The second is the long-tailed distribution of real-world text, where rare but consequential categories such as artistic, inverse, overlapped, occluded, and low-contrast text are under-represented in training data and in existing benchmarks. To address both, the paper proposes Joint-Dataset Learning (JDL), training and testing on the union of many datasets, and introduces the Long-Tailed Benchmark (LTB), 924 images with 2,770 annotated challenging instances grouped into 13 subcategories under intra-instance, inter-instance, and background problem families. If the diagnosis holds, published F-measures from per-dataset fine-tuning overstate generalization, and LTB gives the field a way to measure progress on the cases that actually break detectors in deployment.

What carries the argument

The argument runs on three mechanisms. The Fine-tuning Gap is operationalized by comparing F-measure on a target dataset before and after fine-tuning, and by testing a model fine-tuned on one dataset against another dataset's test set. The Joint-Dataset Learning protocol replaces per-dataset fine-tuning with training on the union $\bigcup_i D_i^{\text{train}}$ and evaluating on $\bigcup_i D_i^{\text{test}}$, applied in the paper to Joint98K, a union of nine datasets with 98,015 training images and 21,675 test images. The Long-Tailed Benchmark is built by Algorithm 1, which keeps ground-truth instances whose maximum Intersection-over-Union with any prediction from four detectors falls below an unreported threshold $t$, plus manual selection from InverseText, Occluded-RoadText, and NightTime-ArT; these instances are organized into three groups (intra-instance, inter-instance, background) and 13 subcategories. Finally, MAEDet pairs a Masked Auto-Encoder backbone with a frozen CLIP attention map that biases reconstruction toward text pixels through a balanced reconstruction loss $\mathcal{L}_{br} = \sum_{i,j} \left(\alpha \mathbf{1}_{M_g > T} + (1-\alpha)\mathbf{1}_{M_g \le T}\right) L_2(I_{ij}, I^{\text{rec}}_{ij})$.

What would settle it

Collect an independent sample of rare-text images without using the four detectors for selection, score the same models on it, and compare the ranking to LTB; if the ranking flips, LTB tracks detector agreement rather than intrinsic difficulty. A cheaper check is to recompute LTB membership across a sweep of the unreported IoU threshold $t$ and see whether the 13-category distribution and model ordering remain stable.

Watch

Extended reading notes

Core claim

The paper's central claim is that the discrepancy between near-saturated academic benchmark results and disappointing real-world behavior in scene text detection has two named causes. The first is the Fine-tuning Gap: under Dataset-Specific Optimization, a model is pretrained broadly and then fine-tuned on one dataset's training split, so its test-set score reflects overfitting to that dataset's biases; the paper reports fine-tuning gains of at least 2.0 F-measure points on the target dataset alongside drops of up to 8.8 points when the same recipe is evaluated on another dataset. The second cause is the long-tailed distribution of text itself: a failure analysis of four detectors on three datasets finds that they miss 7% of text instances, and only 65% of those misses are genuinely hard, with the remainder split between annotation errors and non-Latin scripts; rare categories such as artistic, inverse, overlapped, occluded, and low-contrast text dominate the hard failures. To correct the first cause, the paper advocates Joint-Dataset Learning (JDL), training on the union of datasets and evaluating on the union of test sets, which it shows reduces cross-dataset performance variance. To correct the second, it introduces LTB, a 924-image benchmark whose 2,770 challenging instances are labeled into 13 subcategories under intra-instance, inter-instance, and background problem families, and contributes MAEDet, a self-supervised baseline that reconstructs masked images with a CLIP-derived text-region mask before supervised detection training. If the claim is right, published DSO-era scores overstate generalization and LTB is a diagnostic instrument for the field.

Load-bearing premise

The load-bearing premise is that the text instances missed by the four chosen detectors on three source datasets, plus manually selected hard cases from three other datasets, form a representative sample of the real-world long tail; the unreported IoU threshold $t$ in the filtering algorithm is part of this premise.

Editorial extensions

If this is right

  • Published F-measures from the Dataset-Specific Optimization protocol mix real generalization with benchmark-specific overfitting, so rankings produced under DSO should be re-read as upper bounds rather than deployment predictions.
  • Moving to Joint-Dataset Learning gives a more transferable ranking at no algorithmic cost: the paper reports that cross-dataset variance for DPText-DETR drops from 15.1 to 6.0 F-measure points under JDL.
  • LTB's Hard metric, which scores only the 2,770 challenging instances, separates models on long-tail ability even when their Norm scores over the full annotations look similar.
  • Because the average LTB instance carries about 1.5 challenge attributes, methods tuned to a single failure mode will saturate on isolated benchmarks while still failing on compound cases.
  • Text-aware self-supervised pretraining of the MAEDet type raises detection on tail categories, which suggests that data-side solutions can complement architecture changes in long-tailed text detection.

Reading between the lines

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

  • Extension: because LTB selects failures by consensus of four detectors, an independent test would sample rare-text images with a wholly different detector family; if rankings invert, LTB would be measuring shared blind spots rather than intrinsic long-tail difficulty.
  • Extension: the three-group, 13-subcategory taxonomy is generic enough to transfer to other fine-grained detection tasks, such as traffic-sign or packaging-text detection, where per-dataset fine-tuning similarly hides rare categories; the paper does not make this claim.
  • Extension: publishing the IoU threshold $t$ from Algorithm 1 and a sensitivity curve would let users calibrate how hard LTB is, turning the benchmark into a tunable instrument as detectors improve.
  • Extension: a direct deployment check would be to run cameras or document pipelines on freshly collected scenes, compare failure rates on tail text with LTB Hard scores, and see whether LTB predicts which models fail in the field.
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

5 major / 4 minor

Summary. This paper argues that the gap between high scene-text-detection benchmark scores and poor real-world performance is caused by two factors: a fine-tuning gap from Dataset-Specific Optimization (DSO), and a long-tailed distribution of text instances with rare, challenging categories. To address the first factor, the authors propose Joint-Dataset Learning (JDL), in which models are trained and evaluated on combined training and test sets from multiple datasets. For the second factor, they conduct an error analysis of four detectors on three public test sets and use undetected instances plus manually selected hard cases to build the Long-Tailed Benchmark (LTB), containing 924 images and 2,770 challenging text instances organized into 13 categories. They also propose MAEDet, a self-supervised ViT-based detector with a CLIP-guided reconstruction loss, and report extensive comparisons on Joint98K and LTB. The paper includes code, per-category statistics, and ablations, and it makes a strong causal claim that long-tailed distribution is the primary reason for suboptimal practical performance.

Significance. The paper addresses a real and timely problem: academic benchmarks for scene text detection have become saturated while deployed detectors still fail on hard instances. The JDL protocol is a sensible evaluation alternative and the fine-tuning gap is a plausible phenomenon, supported by Table 1 for two detectors. The LTB benchmark, if constructed reliably, would be a useful diagnostic resource, and the authors provide a concrete taxonomy, explicit per-category counts in Table 7, and a public code repository. The proposal of MAEDet as a self-supervised baseline is also a constructive contribution. However, the significance depends critically on whether LTB truly represents the natural long-tailed distribution rather than the failure modes of the four specific detectors used to filter it; this is exactly where the current manuscript is weakest.

major comments (5)
  1. [Section 4.2 / Algorithm 1] The IoU threshold t in Algorithm 1 is never reported, so the composition of LTB cannot be reproduced. This is not a stylistic omission: t determines whether an instance is 'undetected by all detectors' or merely partially matched, and different choices of t will change both the instance set and the category distribution of LTB. Please report t, justify its value, and provide a sensitivity analysis showing how LTB contents and category counts change over a plausible range of t.
  2. [Section 4.2 / Appendix C.2] LTB instances from ICDAR2015, Total-Text, and ArT are selected as ground-truth boxes whose maximum IoU with predictions from DBNet++, ABCNet v2, DPText-DETR, and SRFormer is below t, and the 13-category taxonomy is derived from these failure-filtered cases. The benchmark therefore measures agreement among the failure modes of those four detectors, not necessarily the natural long-tailed distribution of scene text. The claim that LTB provides 'a credibility benchmark' for real-world detection requires evidence that the selected instances are representative, for example frequency estimates of the 13 categories in unfiltered real-world images, or a demonstration that category distributions are stable across different detector ensembles and thresholds.
  3. [Section 5.2 / Table 4 and Appendix C.1] Joint models are trained on Joint98K, which contains the training splits of ICDAR2015, Total-Text, and ArT, while LTB is constructed from test instances of exactly those same datasets (plus manual additions). Table 4 therefore evaluates joint models on hard instances from datasets whose training distributions were seen during training; this is an in-distribution hard-example test, not an out-of-distribution test of real-world generalization. To support the claim that JDL improves long-tailed robustness, the authors should evaluate LTB on models trained without access to the training splits of the source datasets, or explicitly state the in-distribution nature of the evaluation and its implications.
  4. [Abstract and Section 1] The abstract's causal statement that 'the suboptimal performance in practical settings is primarily attributed to the long-tailed distribution' is not directly evidenced. The paper shows that four detectors miss some instances and that the filtered set has uneven category counts, but it does not measure the frequency of text categories in real-world data, does not estimate category frequencies in training sets, and does not compare the long-tail factor against other potential causes such as domain shift, resolution, or annotation noise. Please soften this to a contributing factor, or add direct evidence such as a training-set frequency analysis and a comparison of failure rates across categories.
  5. [Section 5.3 / Table 5] The hyperparameters of MAEDet (input size, mask threshold T, and balance factor alpha) are selected by directly evaluating on the LTB test set. Tuning the proposed baseline on the benchmark it is meant to support invalidates the reported baseline numbers and creates an unfair baseline for future comparisons. Please use a held-out validation split of LTB for hyperparameter selection, or clearly disclose and justify the test-set tuning.
minor comments (4)
  1. [Appendix C.2 / Table 7] The text says '1,012 occluded text instances (Column 1) and 960 blurred text instances (Column 10)', but Table 7 shows Column 1 is blurred (962 total) and Column 10 is occluded (1,012 total); the column references are reversed.
  2. [Appendix D.2] The sentence 'MAEDet includes tree components' contains a typo; it should read 'three components'.
  3. [Section 4.2 and Appendix C.2] Algorithm 1 and the filtering process appear only in the appendix; the main text refers to the filtering conceptually but does not cite Algorithm 1. Moving the algorithm to the main text or adding an explicit reference would improve readability and reproducibility.
  4. [Appendix B] The subsection 'Similar Protocols in Other Deep Learning Domains' mentions 'Wang et al.' and 'He et al.' without citations; please add the corresponding references.

Circularity Check

1 steps flagged · score 2.0 of 10

LTB's hard split is defined by detector failures, so the same detectors' low LTB scores are partly self-definitional; the independent JDL analysis and the benchmark's usefulness for future methods keep the overall circularity minor.

  1. self definitional [Section 4.2 and Algorithm 1; Table 4]
    "For ICDAR2015, Total-Text, and ArT, we obtain detection results of 4 well-finetuned detectors, as outlined in Table 2. We then develop an algorithm to retain undetected text instances. ... if u < tthen T ′ ← T ′ ∪ g"

    LTB's hard cases are constructed as the ground-truth instances whose maximum IoU with the four chosen detectors is below threshold t. When the paper then reports those same detectors' low F-measures on LTB in Table 4, the 'failure' is partly guaranteed by the benchmark's construction rather than being an independent measurement of long-tailed difficulty. The 13-category taxonomy is derived from this same failure-filtered set, so category prevalence and detector failure rates are not fully independent evidence for the long-tail claim. The fine-tuning gap analysis and the JDL protocol comparison are unaffected by this construction.

full rationale

The paper's central quantitative contributions are empirical comparisons against external annotations: Table 1 directly measures the fine-tuning gap, and Table 3 compares training protocols on standard test sets. These results do not reduce to the paper's definitions or to fitted parameters. The LTB benchmark is a dataset-construction effort, and the detector-assisted screening is disclosed in Algorithm 1; this makes the benchmark reproducible in principle, though the unreported IoU threshold t is a reproducibility weakness rather than a circularity. The self-referential aspect is limited to using the four screening detectors' LTB scores as evidence that detectors struggle with long-tailed categories: since those instances were selected precisely because those detectors miss them, the low scores are partially built in. MAEDet's hyperparameters in Table 5 are tuned on LTB, which weakens the baseline claim but is not a circular derivation of a prediction. No load-bearing self-citations or imported uniqueness theorems were found. Overall circularity is minor and localized, so a score of 2 is appropriate.

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

The central scientific content is empirical, so most load-bearing assumptions are about benchmark representativeness rather than mathematical axioms. The only fitted scalar choices are MAEDet hyperparameters (T, alpha, input size) selected on LTB, plus the unreported IoU threshold t that determines which instances enter LTB. No new physical entities are postulated.

free parameters (4)
  • Mask threshold T in MAEDet = 0.1
    Selected on LTB through ablations in Section 5.3; controls which pixels are treated as text in the reconstruction loss.
  • Balance factor alpha in MAEDet = 0.9
    Selected on LTB through ablations in Section 5.3; weights reconstruction loss between text and background regions.
  • Input image size = 640
    Chosen after ablations on LTB (Exp. 1-3) and used for all MAEDet results, balancing cost and accuracy.
  • IoU threshold t in Algorithm 1 = not reported
    This threshold decides whether a ground-truth instance is considered undetected and thus included in LTB; the value is not given in the paper.
assumptions (3)
  • ad hoc to paper The 13-category taxonomy is complete for long-tailed scene text challenges.
    The taxonomy is introduced by the authors based on qualitative failure analysis in Section 4.1; no inter-annotator agreement or coverage study is reported, yet LTB's validity depends on this categorization.
  • domain assumption Undetected text instances are a valid proxy for real-world long-tailed difficulty.
    Algorithm 1 selects instances missed by four detectors, and the paper assumes these are representative of the long tail; this assumption is not directly validated against real-world failure data.
  • domain assumption Word-level English annotation is the right evaluation scope.
    Non-Latin and line-level annotations are removed in Appendix C.2, so the benchmark covers only English word-level text while the paper generalizes to practical scene text broadly.

how reviews work

0 comments
Cite this review

Pith. "Pith review of The Devil is in Fine-tuning and Long-tailed Problems:A New Benchmark for Scene Text Detection." pith.science (2026). https://pith.science/paper/AQSBO7BX

@misc{pith2026250515649,
  author       = {Pith},
  title        = {Pith review of: The Devil is in Fine-tuning and Long-tailed Problems:A New Benchmark for Scene Text Detection},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/AQSBO7BX}},
  note         = {Machine review of arXiv:2505.15649}
}
read the original abstract

Scene text detection has seen the emergence of high-performing methods that excel on academic benchmarks. However, these detectors often fail to replicate such success in real-world scenarios. We uncover two key factors contributing to this discrepancy through extensive experiments. First, a \textit{Fine-tuning Gap}, where models leverage \textit{Dataset-Specific Optimization} (DSO) paradigm for one domain at the cost of reduced effectiveness in others, leads to inflated performances on academic benchmarks. Second, the suboptimal performance in practical settings is primarily attributed to the long-tailed distribution of texts, where detectors struggle with rare and complex categories as artistic or overlapped text. Given that the DSO paradigm might undermine the generalization ability of models, we advocate for a \textit{Joint-Dataset Learning} (JDL) protocol to alleviate the Fine-tuning Gap. Additionally, an error analysis is conducted to identify three major categories and 13 subcategories of challenges in long-tailed scene text, upon which we propose a Long-Tailed Benchmark (LTB). LTB facilitates a comprehensive evaluation of ability to handle a diverse range of long-tailed challenges. We further introduce MAEDet, a self-supervised learning-based method, as a strong baseline for LTB. The code is available at https://github.com/pd162/LTB.

Figures

Figures reproduced from arXiv: 2505.15649 by the authors.

Figure 1
Figure 1. (a): The performance tendency of scene text detectors in [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. A quantitative analysis of failure cases. The derivation of [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗
Figure 3
Figure 3. Comparison of the Dataset-Specific Optimization (DSO) [PITH_FULL_IMAGE:figures/full_fig_p003_3.png] view at source ↗
Figures from the paper (10 more)
Figure 4
Figure 4. Figure 4: An overview of LTB which collects the challenging issues in scene text detection, involving the [PITH_FULL_IMAGE:figures/full_fig_p004_4.png]
Figure 5
Figure 5. Figure 5: A detailed illustration of MAEDet architecture and training [PITH_FULL_IMAGE:figures/full_fig_p005_5.png]
Figure 6
Figure 6. Figure 6: Detection results of DPText-DETR in practical applica [PITH_FULL_IMAGE:figures/full_fig_p010_6.png]
Figure 7
Figure 7. Figure 7: A visualization of clustering results for images derived [PITH_FULL_IMAGE:figures/full_fig_p010_7.png]
Figure 8
Figure 8. Figure 8: A visualization of the filtering process. Red bounding [PITH_FULL_IMAGE:figures/full_fig_p012_8.png]
Figure 9
Figure 9. Figure 9: Reconstruction results of MAE. with the default ViT feature map of a scale of 1 16 , we produce feature maps of scale { 1 32 , 1 16 , 1 8 , 1 4 } using convolution and deconvolution layers. D.3 Reconstruction Result of MAEDet We visualize the reconstruction results dur…
Figure 10
Figure 10. Figure 10: More examples about the inter-instance challenges of LTB. Red bounding boxes mark challenging text instances and blue ones for [PITH_FULL_IMAGE:figures/full_fig_p015_10.png]
Figure 11
Figure 11. Figure 11: More examples about the inter-instance challenges of LTB. Red bounding boxes mark challenging text instances and blue ones for [PITH_FULL_IMAGE:figures/full_fig_p016_11.png]
Figure 12
Figure 12. Figure 12: More examples about LTB. Purple and brown denote the inter-instance challenges and intra-instance challenges. Red bounding [PITH_FULL_IMAGE:figures/full_fig_p017_12.png]
Figure 13
Figure 13. Figure 13: More examples about the background challenges of LTB. Red bounding boxes mark challenging text instances and blue ones for [PITH_FULL_IMAGE:figures/full_fig_p018_13.png]

Discussion (0). Continue with ORCID 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. Orthogonal Knowledge Refreshing for Domain-Incremental Object Detection

    cs.CV 2026-07 reject novelty 5.0 of 10

    OKR beats exemplar-free domain-incremental detectors by 5.6–6.5 mAP using per-domain LoRA branches plus gradient orthogonality and prototype alignment.

Reference graph

Works this paper leans on

52 extracted references · 50 canonical work pages · cited by 1 Pith paper

  1. [1]

    SRFormer: Text detec- tion transformer with incorporated segmentation and regression

    [Bu et al., 2024] Qingwen Bu, Sungrae Park, Minsoo Khang, and Yichuan Cheng. SRFormer: Text detec- tion transformer with incorporated segmentation and regression. In AAAI, volume 38, pages 855–863,

    work page 2024

  2. [3]

    Total-text: A comprehensive dataset for scene text detection and recognition

    [Ch’ng and Chan, 2017] Chee Kheng Ch’ng and Chee Seng Chan. Total-text: A comprehensive dataset for scene text detection and recognition. In ICDAR, volume 1, pages 935–942. IEEE,

    work page 2017

  3. [9]

    Masked au- toencoders are scalable vision learners

    [He et al., 2022] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Doll´ar, and Ross Girshick. Masked au- toencoders are scalable vision learners. In CVPR, pages 16000–16009,

    work page 2022

  4. [10]

    Reading text in the wild with convolutional neural networks.International journal of computer vision, 116:1–20,

    [Jaderberg et al., 2016] Max Jaderberg, Karen Simonyan, Andrea Vedaldi, and Andrew Zisserman. Reading text in the wild with convolutional neural networks.International journal of computer vision, 116:1–20,

    work page 2016

  5. [11]

    Revisiting scene text recognition: A data perspective

    [Jiang et al., 2023] Qing Jiang, Jiapeng Wang, Dezhi Peng, Chongyu Liu, and Lianwen Jin. Revisiting scene text recognition: A data perspective. In ICCV, pages 20543– 20554,

    work page 2023

  6. [15]

    Textboxes++: A single-shot oriented scene text detec- tor

    [Liao et al., 2018] Minghui Liao, Baoguang Shi, and Xiang Bai. Textboxes++: A single-shot oriented scene text detec- tor. IEEE TIP, 27(8):3676–3690,

    work page 2018

  7. [18]

    ABCNet: Real-time scene text spotting with adaptive bezier-curve network

    [Liu et al., 2020] Yuliang Liu, Hao Chen, Chunhua Shen, Tong He, Lianwen Jin, and Liangwei Wang. ABCNet: Real-time scene text spotting with adaptive bezier-curve network. In CVPR, pages 9809–9818,

    work page 2020

  8. [19]

    Abcnet v2: Adaptive bezier-curve network for real-time end-to- end text spotting

    [Liu et al., 2021] Yuliang Liu, Chunhua Shen, Lianwen Jin, Tong He, Peng Chen, Chongyu Liu, and Hao Chen. Abcnet v2: Adaptive bezier-curve network for real-time end-to- end text spotting. IEEE TPAMI, 44(11):8048–8064,

    work page 2021

Show all 52 references
  1. [20]

    Pbformer: Capturing com- plex scene text shape with polynomial band Transformer

    [Liu et al., 2023] Ruijin Liu, Ning Lu, Dapeng Chen, Cheng Li, Zejian Yuan, and Wei Peng. Pbformer: Capturing com- plex scene text shape with polynomial band Transformer. In ACM MM, pages 2112–2120,

    2023

  2. [21]

    Textsnake: A flexible representation for detecting text of arbitrary shapes

    [Long et al., 2018] Shangbang Long, Jiaqiang Ruan, Wenjie Zhang, Xin He, Wenhao Wu, and Cong Yao. Textsnake: A flexible representation for detecting text of arbitrary shapes. In ECCV, pages 20–36,

    2018

  3. [22]

    Towards end-to-end unified scene text de- tection and layout analysis

    [Long et al., 2022] Shangbang Long, Siyang Qin, Dmitry Panteleev, Alessandro Bissacco, Yasuhisa Fujii, and Michalis Raptis. Towards end-to-end unified scene text de- tection and layout analysis. In CVPR, pages 1049–1059,

    2022

  4. [23]

    Arbitrary reading order scene text spotter with local semantics guidance

    [Lyu et al., 2025] Jiahao Lyu, Wei Wang, Dongbao Yang, Jinwen Zhong, and Yu Zhou. Arbitrary reading order scene text spotter with local semantics guidance. In AAAI, vol- ume 39, pages 5919–5927,

    2025

  5. [25]

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

    [Nayef et al., 2019] Nibal Nayef, Yash Patel, Michal Busta, Pinaki Nath Chowdhury, Dimosthenis Karatzas, Wafa Khlif, Jiri Matas, Umapada Pal, Jean-Christophe Burie, Cheng-lin Liu, et al. Icdar2019 robust reading challenge on multi-lingual scene text detection and recognition—r...

    2019

  6. [26]

    Towards robust real-time scene text detec- tion: From semantic to instance representation learning

    [Qin et al., 2023] Xugong Qin, Pengyuan Lyu, Chengquan Zhang, Yu Zhou, Kun Yao, Peng Zhang, Hailun Lin, and Weiping Wang. Towards robust real-time scene text detec- tion: From semantic to instance representation learning. In ACM MM, pages 2025–2034,

    2023

  7. [27]

    Learning transferable visual models from nat- ural language supervision

    [Radford et al., 2021] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agar- wal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from nat- ural language supervision. In International confere...

    2021

  8. [28]

    Transformer-based text detection in the wild

    [Raisi et al., 2021] Zobeir Raisi, Mohamed A Naiel, Georges Younes, Steven Wardell, and John S Zelek. Transformer-based text detection in the wild. In CVPRW, pages 3162–3171,

    2021

  9. [29]

    Divide rows and conquer cells: Towards structure recognition for large tables

    [Shen et al., 2023] Huawen Shen, Xiang Gao, Jin Wei, Liang Qiao, Yu Zhou, Qiang Li, and Zhanzhan Cheng. Divide rows and conquer cells: Towards structure recognition for large tables. In IJCAI, pages 1369–1377,

    2023

  10. [30]

    Falcon-ui: Understanding gui before following user in- structions

    [Shen et al., 2024] Huawen Shen, Chang Liu, Gengluo Li, Xinlong Wang, Yu Zhou, Can Ma, and Xiangyang Ji. Falcon-ui: Understanding gui before following user in- structions. arXiv preprint arXiv:2412.09362,

    2024 arXiv

  11. [31]

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

    [Singh et al., 2021] Amanpreet Singh, Guan Pang, Mandy Toh, Jing Huang, Wojciech Galuba, and Tal Hassner. Textocr: Towards large-scale end-to-end reasoning for arbitrary-shaped scene text. In CVPR, pages 8802–8812,

    2021

  12. [32]

    Lranet: Towards accurate and efficient scene text detection with low-rank approximation network

    [Su et al., 2024] Yuchen Su, Zhineng Chen, Zhiwen Shao, Yuning Du, Zhilong Ji, Jinfeng Bai, Yong Zhou, and Yu- Gang Jiang. Lranet: Towards accurate and efficient scene text detection with low-rank approximation network. In AAAI, volume 38, pages 4979–4987,

    2024

  13. [33]

    Icdar 2019 competition on large-scale street view text with par- tial labeling-rrc-lsvt

    [Sun et al., 2019] Yipeng Sun, Zihan Ni, Chee-Kheng Chng, Yuliang Liu, Canjie Luo, Chun Chet Ng, Junyu Han, Er- rui Ding, Jingtuo Liu, Dimosthenis Karatzas, et al. Icdar 2019 competition on large-scale street view text with par- tial labeling-rrc-lsvt. In ICDAR, pages 1557–1562. IEEE,

    2019

  14. [34]

    Seglink++: De- tecting dense and arbitrary-shaped scene text by instance- aware component grouping

    [Tang et al., 2019] Jun Tang, Zhibo Yang, Yongpan Wang, Qi Zheng, Yongchao Xu, and Xiang Bai. Seglink++: De- tecting dense and arbitrary-shaped scene text by instance- aware component grouping. PR, 96:106954,

    2019

  15. [35]

    Few could be better than all: Feature sampling and grouping for scene text detection

    [Tang et al., 2022] Jingqun Tang, Wenqing Zhang, Hongye Liu, MingKun Yang, Bo Jiang, Guanglong Hu, and Xiang Bai. Few could be better than all: Feature sampling and grouping for scene text detection. In CVPR, pages 4563– 4572,

    2022

  16. [36]

    Domain adaptive scene text detection via subcategorization

    [Tian et al., 2022] Zichen Tian, Chuhui Xue, Jingyi Zhang, and Shijian Lu. Domain adaptive scene text detection via subcategorization. arXiv,

    2022

  17. [37]

    Efficient and accurate arbitrary-shaped text detection with pixel aggregation network

    [Wang et al., 2019] Wenhai Wang, Enze Xie, Xiaoge Song, Yuhang Zang, Wenjia Wang, Tong Lu, Gang Yu, and Chunhua Shen. Efficient and accurate arbitrary-shaped text detection with pixel aggregation network. In ICCV, pages 8440–8449,

    2019

  18. [38]

    TextRay: Contour-based geometric modeling for arbitrary-shaped scene text detection

    [Wang et al., 2020] Fangfang Wang, Yifeng Chen, Fei Wu, and Xi Li. TextRay: Contour-based geometric modeling for arbitrary-shaped scene text detection. In ACM MM, pages 111–119,

    2020

  19. [39]

    TP- SNet: Reverse thinking of thin plate splines for arbitrary shape scene text representation

    [Wang et al., 2022] Wei Wang, Yu Zhou, Jiahao Lv, Dayan Wu, Guoqing Zhao, Ning Jiang, and Weipinng Wang. TP- SNet: Reverse thinking of thin plate splines for arbitrary shape scene text representation. InACM MM, pages 5014– 5025,

    2022

  20. [40]

    Sequential defor- mation for accurate scene text detection

    [Xiao et al., 2020] Shanyu Xiao, Liangrui Peng, Ruijie Yan, Keyu An, Gang Yao, and Jaesik Min. Sequential defor- mation for accurate scene text detection. In ECCV, pages 108–124. Springer,

    2020

  21. [41]

    Language mat- ters: A weakly supervised vision-language pre-training ap- proach for scene text detection and spotting

    [Xue et al., 2022] Chuhui Xue, Wenqing Zhang, Yu Hao, Shijian Lu, Philip HS Torr, and Song Bai. Language mat- ters: A weakly supervised vision-language pre-training ap- proach for scene text detection and spotting. In ECCV, pages 284–302. Springer,

    2022

  22. [42]

    Textfusenet: Scene text detection with richer fused fea- tures

    [Ye et al., 2020] Jian Ye, Zhe Chen, Juhua Liu, and Bo Du. Textfusenet: Scene text detection with richer fused fea- tures. In IJCAI, pages 516–522. International Joint Con- ferences on Artificial Intelligence Organization,

    2020

  23. [43]

    DPText-DETR: Towards better scene text detection with dynamic points in Transformer

    [Ye et al., 2023] Maoyuan Ye, Jing Zhang, Shanshan Zhao, Juhua Liu, Bo Du, and Dacheng Tao. DPText-DETR: Towards better scene text detection with dynamic points in Transformer. In AAAI, volume 37, pages 3241–3249,

    2023

  24. [44]

    Turning a CLIP model into a scene text detector

    [Yu et al., 2023] Wenwen Yu, Yuliang Liu, Wei Hua, De- qiang Jiang, Bo Ren, and Xiang Bai. Turning a CLIP model into a scene text detector. In CVPR, pages 6978– 6988,

    2023

  25. [45]

    Filling in the blank: Rationale-augmented prompt tuning for textvqa

    [Zeng et al., 2023] Gangyan Zeng, Yuan Zhang, Yu Zhou, Bo Fang, Guoqing Zhao, Xin Wei, and Weiping Wang. Filling in the blank: Rationale-augmented prompt tuning for textvqa. In ACM MM, pages 1261–1272,

    2023

  26. [46]

    Ga-dan: Geometry-aware domain adaptation network for scene text detection and recognition

    [Zhan et al., 2019] Fangneng Zhan, Chuhui Xue, and Shijian Lu. Ga-dan: Geometry-aware domain adaptation network for scene text detection and recognition. In ICCV, pages 9105–9115,

    2019

  27. [47]

    Linguistics-aware masked image modeling for self- supervised scene text recognition

    [Zhang et al., 2025] Yifei Zhang, Chang Liu, Jin Wei, Xi- aomeng Yang, Yu Zhou, Can Ma, and Xiangyang Ji. Linguistics-aware masked image modeling for self- supervised scene text recognition. In CVPR,

    2025

  28. [48]

    EAST: an efficient and accurate scene text detector

    [Zhou et al., 2017] Xinyu Zhou, Cong Yao, He Wen, Yuzhi Wang, Shuchang Zhou, Weiran He, and Jiajun Liang. EAST: an efficient and accurate scene text detector. In CVPR, pages 5551–5560,

    2017

  29. [49]

    Deformable DETR: de- formable transformers for end-to-end object detection

    [Zhu et al., 2021] Xizhou Zhu, Weijie Su, Lewei Lu, Bin Li, Xiaogang Wang, and Jifeng Dai. Deformable DETR: de- formable transformers for end-to-end object detection. In ICLR. OpenReview.net,

    2021

  30. [50]

    reen-fed

    A Powerlessness of Scene Text Detector Real-world scene text is inherently complex, as a single text instance may possess multiple challenging attributes that stump detectors. As illustrated in Figure 6, the text “reen-fed” simultaneously belongs to the occluded, artistic, and...

    2022

  31. [100]

    D.2 Detailed Architecture of MAEDet Like other scene text detectors, MAEDet includes tree com- ponents, backbone, neck, and detection head

    The joint training is conducted with 4 NVIDIA RTX 4090 GPUs. D.2 Detailed Architecture of MAEDet Like other scene text detectors, MAEDet includes tree com- ponents, backbone, neck, and detection head. Backbone We follow the default settings of ViT-B [Dosovitskiy et al., 2021],...

    2021

  32. [640]

    pretraining is conducted with 8 NVIDIA RTX 2080Ti GPUs

    The batch size is set to 16 and the learning rate settings and adjustment scheduler are followed MAE[He et al., 2022]. pretraining is conducted with 8 NVIDIA RTX 2080Ti GPUs. As for the supervised learning stage, we em- ploy an AdamW optimizer with a weight decay of 1e-4. The ...

    2022

  33. [2013]

    Ic- dar 2015 competition on robust reading

    [Karatzas et al., 2015] Dimosthenis Karatzas, Lluis Gomez- Bigorda, Anguelos Nicolaou, Suman Ghosh, Andrew Bag- danov, Masakazu Iwamura, Jiri Matas, Lukas Neumann, Vijay Ramaseshan Chandrasekhar, Shijian Lu, et al. Ic- dar 2015 competition on robust reading. In ICDAR, pages 11...

    2015

  34. [2015]

    Textboxes: A fast text detector with a single deep neural network

    [Liao et al., 2017] Minghui Liao, Baoguang Shi, Xiang Bai, Xinggang Wang, and Wenyu Liu. Textboxes: A fast text detector with a single deep neural network. In AAAI, vol- ume 31,

    2017

  35. [2016]

    Mask R-CNN

    [He et al., 2017] Kaiming He, Georgia Gkioxari, Piotr Doll´ar, and Ross Girshick. Mask R-CNN. In ICCV, pages 2961–2969,

    2017

  36. [2017]

    Ic- dar2019 robust reading challenge on arbitrary-shaped text- rrc-art

    [Chng et al., 2019] Chee Kheng Chng, Yuliang Liu, Yipeng Sun, Chun Chet Ng, Canjie Luo, Zihan Ni, ChuanMing Fang, Shuaitao Zhang, Junyu Han, Errui Ding, et al. Ic- dar2019 robust reading challenge on arbitrary-shaped text- rrc-art. In ICDAR, pages 1571–1576. IEEE,

    2019

  37. [2018]

    Real-time scene text detec- tion with differentiable binarization

    [Liao et al., 2020] Minghui Liao, Zhaoyi Wan, Cong Yao, Kai Chen, and Xiang Bai. Real-time scene text detec- tion with differentiable binarization. In AAAI, volume 34, pages 11474–11481,

    2020

  38. [2019]

    An image is worth 16x16 words: Transformers for image recognition at scale

    [Dosovitskiy et al., 2021] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Min- derer, Georg Heigold, and Sylvain Gelly. An image is worth 16x16 words: Transformers for image recognition at s...

    2021

  39. [2020]

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

    [Liu et al., 2017] Yuliang Liu, Lianwen Jin, Shuaitao Zhang, and Sheng Zhang. Detecting curve text in the wild: New dataset and new solution. arXiv,

    2017

  40. [2021]

    I3CL: Intra-and inter-instance collaborative learning for arbitrary-shaped scene text detection

    [Du et al., 2022] Bo Du, Jian Ye, Jing Zhang, Juhua Liu, and Dacheng Tao. I3CL: Intra-and inter-instance collaborative learning for arbitrary-shaped scene text detection. IJCV, 130(8):1961–1977,

    2022

  41. [2022]

    Synthetic data for text localization in nat- ural images

    [Gupta et al., 2016] Ankush Gupta, Andrea Vedaldi, and An- drew Zisserman. Synthetic data for text localization in nat- ural images. In CVPR, pages 2315–2324,

    2016

  42. [2023]

    Icdar 2013 robust reading competition

    [Karatzas et al., 2013] Dimosthenis Karatzas, Faisal Shafait, Seiichi Uchida, Masakazu Iwamura, Lluis Gomez i Big- orda, Sergi Robles Mestre, Joan Mas, David Fernan- dez Mota, Jon Almazan Almazan, and Lluis Pere De Las Heras. Icdar 2013 robust reading competition. In ICDAR, pa...

    2013

  43. [2024]

    Textdiffuser: Diffusion models as text painters

    [Chen et al., 2024] Jingye Chen, Yupan Huang, Tengchao Lv, Lei Cui, Qifeng Chen, and Furu Wei. Textdiffuser: Diffusion models as text painters. NeurIPS, 36,

    2024

  44. [2025]

    Icdar2017 robust reading challenge on multi-lingual scene text detection and script identification-rrc-mlt

    [Nayef et al., 2017] Nibal Nayef, Fei Yin, Imen Bizid, Hyunsoo Choi, Yuan Feng, Dimosthenis Karatzas, Zhenbo Luo, Umapada Pal, Christophe Rigaud, Joseph Chazalon, et al. Icdar2017 robust reading challenge on multi-lingual scene text detection and script identification-rrc-mlt....

    2017

Pith tools

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