Pith. sign in

REVIEW 3 major objections 6 minor 33 references

EdgeSpotter: Multi-Scale Dense Text Spotting for Industrial Panel Monitoring

T0 review · 3 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 EdgeSpotter claims to outperform eight state-of-the-art text spotting methods on industrial panel monitoring, reaching detection F1 of 96.31% and recognition H of 81.25% on the new IPM validation set.

desk verdict A useful industrial text spotting dataset and a modest architecture tweak, undermined by an evaluation that relies entirely on a private validation split with no error bars. read the letter →

arxiv 2506.07112 v1 pith:EISAVWSA submitted 2025-06-08 cs.CV

Changhong Fu , Hua Lin , Haobo Zuo , Liangliang Yao , Liguo Zhang This is my paper
classification cs.CV
keywords textspottingindustrialpanelmonitoringmulti-scalefeaturefusionCatmull-RomsplinesefficientmixerTransformerdenseedgeAIIPMbenchmark
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

EdgeSpotter is proposed as an end-to-end text spotter built for industrial panel monitoring, a setting where text is often dense, highly variable in scale, and lit unevenly. The paper tries to establish that combining a linear-complexity multi-level feature mixer with Catmull-Rom spline feature sampling lets a single lightweight network outperform eight existing text spotting methods on a new industrial panel benchmark, while still running fast enough for edge deployment. The reported results support this: on the IPM validation set EdgeSpotter reaches detection F1 of 96.31% and recognition H of 81.25%, surpassing the second-best methods by 0.63 and 3.32 points respectively. The authors also introduce the IPM dataset and show real-world tests on a self-built edge vision system at over 25 FPS.

What carries the argument

The two load-bearing components are an encoder with efficient mixer Transformer (EMT) and feature sampling with Catmull-Rom splines (FSCRS). EMT concatenates the last three backbone feature maps with a downsampled sixth level into one token sequence, then applies a linear-complexity additive attention block: a learnable vector $W_m$ weights the query attention across levels, and element-wise multiplication keeps cost linear instead of quadratic. FSCRS predicts four Catmull-Rom control points per feature pixel with an MLP, selects the top-$K$ proposals by score ($K=100$), and samples $n=25$ points along each spline to form positional queries for the Transformer decoder; four prediction heads then emit instance class, character class, center curve points, and bounding boxes. The spline sampler is the mechanism that directly encodes text shape, position, and semantics, while the mixer supplies cross-scale spatial context.

What would settle it

Release the IPM validation set and run EdgeSpotter's code on it exactly as described; if the reported 96.31% F1 and 81.25% H do not reproduce, or the margin over ABCNet v2 and DeepSolo disappears under the same evaluation protocol, the central claim is falsified. A stronger test would hold out panels from device models and lighting conditions not seen in training.

Watch

Extended reading notes

Core claim

The paper's central claim is that a text spotter can be made accurate and efficient in dense, multi-scale industrial panel images by explicitly modelling text shape with Catmull-Rom splines and by fusing multi-level backbone features with an efficient attention mechanism. The evidence is the IPM validation set, where EdgeSpotter reports 98.16% precision, 94.52% recall (F1 96.31%) for detection and 82.53% precision, 80.02% recall (H 81.25%) for end-to-end recognition, against eight published methods. Ablation experiments attribute the gains to both new modules: the spline-based feature sampling raises F1 by 5.43 points over the baseline and the efficient mixer raises F1 by 4.77 points, and together they improve recognition H by nearly 7 points. The paper further claims that the method generalizes across seven challenge attributes—side view, strong light, reflection, shadow, special characters, multi-scale, and dense characters—and that real deployments on panel devices maintain accuracy above 25 FPS.

Load-bearing premise

The load-bearing premise is that the IPM validation set represents real industrial panel monitoring and was not used to tune the method; because that dataset is not yet public, no independent researcher can check this.

Editorial extensions

If this is right

  • If the reported performance holds, manual periodic reading of industrial panels can be replaced by continuous 24/7 automated spotting, capturing full lifecycle data without high-intensity human labor.
  • At over 25 FPS on an Intel i9 plus NVIDIA RTX 3070 system, the method fits real-time edge monitoring rather than requiring server-side processing.
  • The IPM benchmark with seven challenge attributes gives industrial text spotting a common testbed, so future methods can be compared on dense text, multi-scale text, and lighting challenges.
  • If EdgeSpotter's gains are real, the same architecture should transfer to related dense-text devices such as meters, oscilloscopes, and temperature controllers without redesign.

Reading between the lines

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

  • The paper does not evaluate on public scene-text benchmarks, so whether these gains generalize beyond the IPM domain is untested; a clean extension would be to run the released model on ICDAR 2015 and CTW1500 and compare with the same baselines.
  • The reported top-$K$ observation—lower-scoring text instances receive more control points—suggests spline-point density could be used as a confidence calibration signal, but the paper does not pursue that direction.
  • Because IPM is not yet public and all hyperparameters were chosen on its validation split, the headline numbers are internal-benchmark claims until an independent replication is possible.
  • The real-world demo reports frame rate and accuracy but not power draw or latency breakdown; an extension would be to measure energy per frame on the edge system.
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

3 major / 6 minor

Summary. The paper proposes EdgeSpotter, an end-to-end Transformer-based text spotter for industrial panel monitoring. The architecture combines an efficient-mixer Transformer encoder that fuses multi-level backbone features with a learned scale-adaptive vector, and a feature sampling module (FSCRS) that uses Catmull-Rom splines to generate curve proposals for text instances. The method is evaluated on a new private dataset, IPM, with 1,200 training and 300 validation images, and compared with eight text-spotting baselines. The paper reports superior detection and recognition performance (F1 = 96.31%, H = 81.25%), ablation results for the two proposed modules, and a real-world deployment demonstration on a self-built edge vision system.

Significance. If the reported performance is reliable, the paper would offer a practical lightweight solution for an underexplored industrial application and a reasonable technical novelty in combining multi-level feature mixing with spline-based feature sampling. The authors provide a new task-specific dataset, qualitative comparisons, and a deployment study, and they promise to release code; these are strengths. However, the empirical evidence is not yet sufficient: all quantitative results come from the authors' own validation split, the training budgets of the baselines are not uniform as claimed, and the small reported margins are not accompanied by error bars or statistical tests. The architectural ideas are plausible and the paper is generally readable, so the issues are addressable with additional experiments rather than being fundamental flaws.

major comments (3)
  1. [§IV-A, §IV-C.1, Table I] The central superiority claim is not supported by the evaluation protocol. The paper defines 1,200 training and 300 validation images for IPM but never introduces a held-out test split, and all results in Table I, Fig. 6, and Table II are reported on this validation set. The hyperparameters K=100, n=25, and τ=0.5 are set empirically on the same validation set (§III-B and §IV-B.1), creating a selection loop. The reported margins over the second-best method are +0.63 F1 and +3.32 H, but no confidence intervals, repeated runs, or significance tests are provided. To establish the claimed superiority, please add a held-out test set (or repeated cross-validation) with mean±std over multiple seeds, and describe the hyperparameter selection procedure so that the validation set is not used both for tuning and for final reporting.
  2. [§IV-B.1] The statement that all models were trained for a uniform 12k iterations is directly contradicted by the parenthetical note that cascade-based methods may be trained for more iterations to ensure results. The paper does not report per-baseline iteration counts, learning-rate schedules, or convergence curves, and the comparison in Table I may therefore reflect undertrained baselines rather than genuine methodological advantage. Please report the exact training configuration for every baseline, show convergence behavior, and either enforce truly uniform budgets or justify different budgets with evidence that each baseline has converged.
  3. [§IV-D, Table II] The ablation description reverses the roles of EMT and FSCRS. Under the natural reading of the table columns (EMT and FSCRS), the row with F1=95.49% and H=79.04% is the EMT-only configuration, while the row with F1=94.83% and H=76.98% is the FSCRS-only configuration. The text, however, states that after introducing FSCRS, F1 increases to 95.49%, and that introduction of EMT alone results in a 4.77% increase in F1 and a 2.64% increase in H, which swaps the two single-module results. In addition, the module-presence symbols are rendered as exclamation marks in the PDF, making the rows ambiguous. Please correct the text, use unambiguous checkmarks or labels, and update the attribution of each gain.
minor comments (6)
  1. [§III-A, Eq. (1)] Equation (1) contains an unmatched closing parenthesis; it should read X = EM(LN(F)) + F. Please also clarify the broadcasting of the attention vector Wattn in Eq. (4).
  2. [§III-B, Eq. (5)] The notation σ^{-1} is used without defining the inverse sigmoid; please define it explicitly.
  3. [§IV-A] The dataset description says the remaining images are classified according to various challenge attributes but does not report how many images belong to each attribute or whether the attributes overlap; please provide these counts.
  4. [§IV-B.1] Please specify the input image resolution and batch size used for all experiments, and fix the grammatical error in the sentence that says EdgeSpotter predict 96 classes.
  5. [Table I] The method name Mask TexSpotter v3 should be Mask TextSpotter v3, and the caption claims red represents the best result, but no red highlighting is visible in the table.
  6. [§IV-E, Fig. 7] The average score for each test group is not defined; please state what score is being averaged and how the ground truth is established in the deployment tests. The speed claim of 25 FPS should also state the input resolution and whether it covers the full end-to-end pipeline on the edge device.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the paper is an empirical benchmark study with no derivation chain that reduces to its own inputs.

full rationale

EdgeSpotter is an applied computer-vision paper whose claims are empirical measurements on a newly constructed benchmark, not consequences of a formal derivation. The two proposed components, EMT (Eqs. 1-4) and FSCRS (Eqs. 5-8), are defined as network architectures and are evaluated by ablations; their reported contributions are measured improvements, not quantities that were inserted into the ablations by construction. The comparison with eight external baselines (Section IV-C) uses published methods, and the paper contains no self-citations or self-authored uniqueness theorems that could be load-bearing. The only mild concern is that hyperparameters such as tau=0.5, K=100, and n=25 are described as set empirically and all results, including baseline comparisons, are reported on the same private IPM validation set, so there is a risk of validation-set overfitting and of uneven baseline training budgets despite the claim of uniform 12k iterations. That is an experimental-validity and correctness concern, not circularity, because the reported F1 and H scores are not logically forced to equal any input by definition. Accordingly, the paper is self-contained with respect to its empirical claims, and no circular step can be exhibited from the text.

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

The central claim rests on the representativeness of the IPM dataset, standard deep-learning training assumptions, and the choice of Catmull-Rom splines for panel text. No new physical entities are introduced; the learnable vector Wm is a model parameter within an existing architecture, not an independent postulate requiring external evidence.

free parameters (4)
  • Catmull-Rom tension τ = 0.5
    Set empirically in Section III.B to control spline smoothness; no sensitivity analysis is provided to justify the choice.
  • Number of proposals K = 100
    Hyperparameter chosen in Section IV.B; not ablated or justified relative to other values.
  • Sampled points per curve n = 25
    Hyperparameter chosen in Section IV.B; no analysis of its effect on text recognition accuracy.
  • Training iterations = 12k
    Uniform iteration count stated in Section IV.B.1; chosen to compare methods but not validated as sufficient or fair for all baselines.
assumptions (5)
  • domain assumption Catmull-Rom spline interpolation accurately models the shape of panel text.
    Invoked in Section III.B (FSCRS) and required for the sampled positional queries to represent text instances; no independent evidence in the paper shows Catmull-Rom is superior to Bezier for this domain.
  • domain assumption The IPM dataset, especially its validation split, is representative of industrial panel monitoring and large enough to rank methods.
    All quantitative claims rely on the IPM validation set (Section IV.C), which the authors constructed; representativeness is asserted but not independently verified.
  • standard math The efficient mixer formulation (Wattn = K Wm, EM = phi(phi(V * Wattn / sqrt(D)) + Q)) captures useful multi-scale attention with linear complexity.
    Based on SwiftFormer [29], a published architecture; the paper assumes this attention variant improves multi-scale feature fusion without proof beyond its empirical results.
  • domain assumption Backbone features from the last three stages of ResNet50 are sufficient for multi-scale text spotting.
    Assumed in Section III.A; typical for text spotters, but not justified for this specific domain or compared against other backbone feature choices.
  • domain assumption Uniform training of all compared methods for 12k iterations with default hyperparameters is a fair comparison.
    Stated in Section IV.B.1; if baselines were not tuned, the comparison may be biased in favor of the proposed method, which was developed with the authors' tuning process.

how reviews work

0 comments
Cite this review

Pith. "Pith review of EdgeSpotter: Multi-Scale Dense Text Spotting for Industrial Panel Monitoring." pith.science (2026). https://pith.science/paper/EISAVWSA

@misc{pith2026250607112,
  author       = {Pith},
  title        = {Pith review of: EdgeSpotter: Multi-Scale Dense Text Spotting for Industrial Panel Monitoring},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/EISAVWSA}},
  note         = {Machine review of arXiv:2506.07112}
}
read the original abstract

Text spotting for industrial panels is a key task for intelligent monitoring. However, achieving efficient and accurate text spotting for complex industrial panels remains challenging due to issues such as cross-scale localization and ambiguous boundaries in dense text regions. Moreover, most existing methods primarily focus on representing a single text shape, neglecting a comprehensive exploration of multi-scale feature information across different texts. To address these issues, this work proposes a novel multi-scale dense text spotter for edge AI-based vision system (EdgeSpotter) to achieve accurate and robust industrial panel monitoring. Specifically, a novel Transformer with efficient mixer is developed to learn the interdependencies among multi-level features, integrating multi-layer spatial and semantic cues. In addition, a new feature sampling with catmull-rom splines is designed, which explicitly encodes the shape, position, and semantic information of text, thereby alleviating missed detections and reducing recognition errors caused by multi-scale or dense text regions. Furthermore, a new benchmark dataset for industrial panel monitoring (IPM) is constructed. Extensive qualitative and quantitative evaluations on this challenging benchmark dataset validate the superior performance of the proposed method in different challenging panel monitoring tasks. Finally, practical tests based on the self-designed edge AI-based vision system demonstrate the practicality of the method. The code and demo will be available at https://github.com/vision4robotics/EdgeSpotter.

Figures

Figures reproduced from arXiv: 2506.07112 by the authors.

Figure 1
Figure 1. An overview of state-of-the-art (SOTA) text spotting methods [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. An overview of the proposed EdgeSpotter. The modules from the left to right are [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Detailed workflow of efficient mixer. The input features are [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (4 more)
Figure 4
Figure 4. Figure 4: Density maps of control points for Top-K scores. Text instances [PITH_FULL_IMAGE:figures/full_fig_p005_4.png]
Figure 5
Figure 5. Figure 5: Comparison of EdgeSpotter spotting results with other SOTA spotting results. Areas with incorrect identification ( [PITH_FULL_IMAGE:figures/full_fig_p006_5.png]
Figure 7
Figure 7. Figure 7: Visualization of real-world tests. In real-world applications, scores above the [PITH_FULL_IMAGE:figures/full_fig_p007_7.png]
Figure 6
Figure 6. Figure 6: The spotting accuracy of EdgeSpotter and other SOTA methods [PITH_FULL_IMAGE:figures/full_fig_p007_6.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

33 extracted references · 32 canonical work pages

  1. [1]

    A Review on Recent Advances in Vision-Based Defect Recognition towards Industrial Intelligence,

    Y . Gao, X. Li, X. V . Wang, L. Wang, and L. Gao, “A Review on Recent Advances in Vision-Based Defect Recognition towards Industrial Intelligence,” Journal of Manufacturing Systems , vol. 62, pp. 753–766, 2022

    work page 2022

  2. [2]

    Methods for Monitoring the Photovoltaic Panel: A Review,

    Y . Shang, H. He, Z. Zhang, X. Zhang, Z. Shen, J. Yin, X. Miao et al., “Methods for Monitoring the Photovoltaic Panel: A Review,” in Proceedings of the International Conference on Agro-Geoinformatics (Agro-Geoinformatics), 2024, pp. 1–5

    work page 2024

  3. [3]

    A Comprehensive Review of Energy Storage Technology Development and Application for Pure Electric Vehicles,

    F. Jiang, X. Yuan, L. Hu, G. Xie, Z. Zhang, X. Li, J. Hu, C. Wang, and H. Wang, “A Comprehensive Review of Energy Storage Technology Development and Application for Pure Electric Vehicles,” Journal of Energy Storage, vol. 86, p. 111159, 2024

    work page 2024

  4. [4]

    An End-to-End Tag Recognition Architecture for Industrial Meter,

    X. Deng, X. Chen, D. Cao, K. Ren, and P. Z. Sun, “An End-to-End Tag Recognition Architecture for Industrial Meter,”IEEE Transactions on Industrial Informatics , vol. 20, no. 1, pp. 117–126, 2023

    work page 2023

  5. [5]

    OCR- Diff: A Two-Stage Deep Learning Framework for Optical Character Recognition Using Diffusion Model in Industrial Internet-of-Things,

    C.-W. Park, V . Palakonda, S. Yun, I.-M. Kim, and J.-M. Kang, “OCR- Diff: A Two-Stage Deep Learning Framework for Optical Character Recognition Using Diffusion Model in Industrial Internet-of-Things,” IEEE Internet of Things Journal , vol. 11, no. 15, pp. 25 997–26 000, 2024

    work page 2024

  6. [6]

    PAN++: Towards Efficient and Accurate End-to-End Spotting of Arbitrarily-Shaped Text,

    W. Wang, E. Xie, X. Li, X. Liu, D. Liang, Z. Yang, T. Lu, and C. Shen, “PAN++: Towards Efficient and Accurate End-to-End Spotting of Arbitrarily-Shaped Text,” IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 44, no. 9, pp. 5349–5367, 2021

    work page 2021

  7. [7]

    ABCNet: Real- Time Scene Text Spotting with Adaptive Bezier-Curve Network,

    Y . Liu, H. Chen, C. Shen, T. He, L. Jin, and L. Wang, “ABCNet: Real- Time Scene Text Spotting with Adaptive Bezier-Curve Network,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020, pp. 9809–9818

    work page 2020

  8. [8]

    ABCNet v2: Adaptive Bezier-Curve Network for Real-Time End- to-End Text Spotting,

    Y . Liu, C. Shen, L. Jin, T. He, P. Chen, C. Liu, and H. Chen, “ABCNet v2: Adaptive Bezier-Curve Network for Real-Time End- to-End Text Spotting,” IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 44, no. 11, pp. 8048–8064, 2021

    work page 2021

Show all 33 references
  1. [9]

    Attention Is All You Need,

    A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. N. Gomez, Ł. Kaiser, and I. Polosukhin, “Attention Is All You Need,” in Proceedings of Neural Information Processing Systems (NIPS) , vol. 30, 2017, pp. 20 750—-20 762

    2017

  2. [10]

    A Survey on Vision Transformer,

    K. Han, Y . Wang, H. Chen, X. Chen, J. Guo, Z. Liu, Y . Tang, A. Xiao, C. Xu, Y . Xuet al., “A Survey on Vision Transformer,” IEEE Transactions on Pattern Analysis and Machine Intelligence , vol. 45, no. 1, pp. 87–110, 2022

    2022

  3. [11]

    Swintextspotter: Scene Text Spotting via Better Synergy between Text Detection and Text Recognition,

    M. Huang, Y . Liu, Z. Peng, C. Liu, D. Lin, S. Zhu, N. Yuan, K. Ding, and L. Jin, “Swintextspotter: Scene Text Spotting via Better Synergy between Text Detection and Text Recognition,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR),...

    2022

  4. [12]

    Deepsolo: Let Transformer Decoder with Explicit Points Solo for Text Spotting,

    M. Ye, J. Zhang, S. Zhao, J. Liu, T. Liu, B. Du, and D. Tao, “Deepsolo: Let Transformer Decoder with Explicit Points Solo for Text Spotting,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023, pp. 19 348–19 357

    2023

  5. [13]

    DNTextSpotter: Arbitrary-Shaped Scene Text Spotting via Improved Denoising Training,

    Q. Qiao, Y . Xie, J. Gao, T. Wu, S. Huang, J. Fan, Z. Cao, Z. Wang, and Y . Zhang, “DNTextSpotter: Arbitrary-Shaped Scene Text Spotting via Improved Denoising Training,” in Proceedings of the ACM Inter- national Conference on Multimedia (ACM MM) , 2024, pp. 10 134– 10 143

    2024

  6. [14]

    Faster R-CNN: Towards Real-Time Object Detection with Region Proposal Networks,

    S. Ren, K. He, R. Girshick, and J. Sun, “Faster R-CNN: Towards Real-Time Object Detection with Region Proposal Networks,” IEEE Transactions on Pattern Analysis and Machine Intelligence , vol. 39, no. 6, pp. 1137–1149, 2016

    2016

  7. [15]

    EAST: an Efficient and Accurate Scene Text Detector,

    X. Zhou, C. Yao, H. Wen, Y . Wang, S. Zhou, W. He, and J. Liang, “EAST: an Efficient and Accurate Scene Text Detector,” in Pro- ceedings of the IEEE conference on Computer Vision and Pattern Recognition (CVPR), 2017, pp. 5551–5560

    2017

  8. [16]

    Real-Time Scene Text Detection with Differentiable Binarization,

    M. Liao, Z. Wan, C. Yao, K. Chen, and X. Bai, “Real-Time Scene Text Detection with Differentiable Binarization,” in Proceedings of the AAAI Conference on Artificial Intelligence (AAAI) , vol. 34, no. 07, 2020, pp. 11 474–11 481

    2020

  9. [17]

    AON: Towards Arbitrarily-Oriented Text Recognition,

    Z. Cheng, Y . Xu, F. Bai, Y . Niu, S. Pu, and S. Zhou, “AON: Towards Arbitrarily-Oriented Text Recognition,” in Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) , 2018, pp. 5571–5579

    2018

  10. [18]

    Towards End-to-End Text Spotting with Convolutional Recurrent Neural Networks,

    H. Li, P. Wang, and C. Shen, “Towards End-to-End Text Spotting with Convolutional Recurrent Neural Networks,” in Proceedings of the IEEE International Conference on Computer Vision (ICCV), 2017, pp. 5238–5246

    2017

  11. [19]

    An End-to-End Textspotter with Explicit Alignment and Attention,

    T. He, Z. Tian, W. Huang, C. Shen, Y . Qiao, and C. Sun, “An End-to-End Textspotter with Explicit Alignment and Attention,” in Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2018, pp. 5020–5029

    2018

  12. [20]

    Mask Textspotter: An End-to-End Trainable Neural Network for Spotting Text with Arbitrary Shapes,

    P. Lyu, M. Liao, C. Yao, W. Wu, and X. Bai, “Mask Textspotter: An End-to-End Trainable Neural Network for Spotting Text with Arbitrary Shapes,” in Proceedings of the European Conference on Computer Vision (ECCV), 2018, pp. 67–83

    2018

  13. [21]

    Textdragon: An End-to-End Framework for Arbitrary Shaped Text Spotting,

    W. Feng, W. He, F. Yin, X.-Y . Zhang, and C.-L. Liu, “Textdragon: An End-to-End Framework for Arbitrary Shaped Text Spotting,” in Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2019, pp. 9076–9085

    2019

  14. [22]

    Mask TextSpotter v3: Segmentation Proposal Network for Robust Scene Text Spotting,

    M. Liao, G. Pang, J. Huang, T. Hassner, and X. Bai, “Mask TextSpotter v3: Segmentation Proposal Network for Robust Scene Text Spotting,” in Proceedings of the European Conference on Computer Vision (ECCV), 2020, pp. 706–722

    2020

  15. [23]

    Mango: A Mask Attention Guided One-Stage Scene Text Spotter,

    L. Qiao, Y . Chen, Z. Cheng, Y . Xu, Y . Niu, S. Pu, and F. Wu, “Mango: A Mask Attention Guided One-Stage Scene Text Spotter,” in Proceedings of the AAAI Conference on Artificial Intelligence (AAAI) , vol. 35, no. 3, 2021, pp. 2467–2476

    2021

  16. [24]

    Text Spotting Transformers,

    X. Zhang, Y . Su, S. Tripathi, and Z. Tu, “Text Spotting Transformers,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022, pp. 9519–9528

    2022

  17. [25]

    Fast R-CNN,

    R. Girshick, “Fast R-CNN,” in Proceedings of the IEEE International Conference on Computer Vision (ICCV) , 2015, pp. 1440–1448

    2015

  18. [26]

    Scatter: Selective Context Attentional Scene Text Recognizer,

    R. Litman, O. Anschel, S. Tsiper, R. Litman, S. Mazor, and R. Man- matha, “Scatter: Selective Context Attentional Scene Text Recognizer,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020, pp. 11 962–11 972

    2020

  19. [27]

    Mask R-CNN,

    K. He, G. Gkioxari, P. Doll ´ar, and R. Girshick, “Mask R-CNN,” in Proceedings of the IEEE International Conference on Computer Vision (ICCV), 2017, pp. 2961–2969

    2017

  20. [28]

    Glass: Global to Local Attention for Scene-Text Spotting,

    R. Ronen, S. Tsiper, O. Anschel, I. Lavi, A. Markovitz, and R. Man- matha, “Glass: Global to Local Attention for Scene-Text Spotting,” in Procedings of the European Conference on Computer Vision (ECCV) , 2022, pp. 249–266

    2022

  21. [29]

    Swiftformer: Efficient Additive Attention for Transformer- based Real-Time Mobile Vision Applications,

    A. Shaker, M. Maaz, H. Rasheed, S. Khan, M.-H. Yang, and F. S. Khan, “Swiftformer: Efficient Additive Attention for Transformer- based Real-Time Mobile Vision Applications,” in Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) , 2023, pp. 17 425–17 436

    2023

  22. [30]

    Catmull-Rom Splines,

    C. Twigg, “Catmull-Rom Splines,” Computer, vol. 41, no. 6, pp. 4–6, 2003

    2003

  23. [31]

    PGNet: Real-Time Arbitrarily-Shaped Text Spotting with Point Gathering Network,

    P. Wang, C. Zhang, F. Qi, S. Liu, X. Zhang, P. Lyu, J. Han, J. Liu, E. Ding, and G. Shi, “PGNet: Real-Time Arbitrarily-Shaped Text Spotting with Point Gathering Network,” in Proceedings of the AAAI Conference on Artificial Intelligence (AAAI) , vol. 35, no. 4, 2021, pp. 2782–2790

    2021

  24. [32]

    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. Luet al., “ICDAR 2015 Competition on Robust Reading,” in Proceedings of the International Conference on Document Analysis and Recognition (ICDAR), 2015...

    2015

  25. [33]

    Curved Scene Text Detection via Transverse and Longitudinal Sequence Connection,

    Y . Liu, L. Jin, S. Zhang, C. Luo, and S. Zhang, “Curved Scene Text Detection via Transverse and Longitudinal Sequence Connection,” Pattern Recognition, vol. 90, pp. 337–345, 2019

    2019

Pith tools

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