Pith. sign in

REVIEW 4 major objections 6 minor 74 references

MetaWriter: Personalized Handwritten Text Recognition Using Meta-Learned Prompt Tuning

T0 review · 4 major / 6 minor · reviewed 2026-08-07 · deepseek-v4-flash

Pith's one-line read The paper shows that a handwriting recognizer can be personalized to a new writer by tuning only tiny prompt vectors, using unlabeled images and a meta-learned initialization.

desk verdict A real step for unlabeled writer adaptation, but the missing support-set selection rule and inconsistent ablation numbers keep the SOTA claim from being fully reproducible. read the letter →

arxiv 2505.20513 v1 pith:RTCUVKJZ submitted 2025-05-26 cs.CV

classification cs.CV
keywords handwrittentextrecognitionwriterpersonalizationprompttuningmeta-learningmaskedautoencodertest-timeadaptationself-supervisedlearningparameter-efficientfine-tuning
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 asserts that a handwriting recognizer can be personalized to an unseen writer by tuning only tiny prompt vectors—padding values inserted around the input images—using a masked-autoencoder reconstruction objective, with no labeled examples at test time. Meta-learning across training writers sets the initial values of these prompts so that a single gradient step on the structural-similarity (SSIM) reconstruction loss also lowers transcription error. On the IAM line-level benchmark, this reaches 3.36% character error and 10.32% word error with five unlabeled lines per writer, while updating 0.08M parameters, and similar gains appear on RIMES. The practical point is that writer adaptation could run on-device and annotation-free, avoiding the user-transcription burden of earlier approaches.

What carries the argument

The load-bearing mechanism is the bi-level optimization of meta-prompts. In the inner loop, writer-specific prompts are obtained from a shared initialization by one gradient step on the masked-autoencoder reconstruction loss $L_{ada}=1-\mathrm{SSIM}(x,\hat{x})$, using a few unlabeled images of that writer. In the outer loop, the initialization is updated with the cross-entropy text-prediction loss $L_{pred}$ computed on a disjoint labeled query set from the same training writers. The prompts are implemented as tunable padding values inserted into the first 18 convolutional layers of the frozen encoder, so adaptation touches roughly 1% of the model's parameters.

What would settle it

For a held-out set of writers, run the adaptation update while tracking both the SSIM reconstruction loss and the CER on a labeled calibration set; if the reconstruction loss drops while CER rises or stays flat for multiple writers, the alignment premise is violated. Equivalently, replace the meta-learned prompt initialization with a random initialization and repeat the same 5-shot unlabeled adaptation—if the random prompts match the meta-learned ones in CER, the meta-learning component contributes nothing.

Watch

Extended reading notes

Core claim

The central discovery is that writer-specific style can be captured in a small set of visual prompts rather than in the network weights. Prompt vectors are placed as learnable padding in the first 18 convolutional layers of a fixed image encoder; at test time only these prompts are updated, by minimizing a self-supervised MAE reconstruction loss (1 − SSIM) on a few masked unlabeled images. Because this reconstruction gradient is not by itself guaranteed to reduce recognition error, the paper uses meta-learning to optimize the prompt initialization over training writers, so the inner-loop update transfers to unseen writers. The paper reports that this recipe outperforms prior instance-specific and writer-specific HTR methods on IAM and RIMES, reaching a 3.36% CER and 10.32% WER on IAM with five-shot unlabeled adaptation.

Load-bearing premise

The load-bearing premise is that a single gradient step on the image-reconstruction loss, computed from a few unlabeled images of a new writer, moves the prompts in a direction that also lowers text recognition error; nothing guarantees this alignment, and meta-learning can only make it hold on average for training writers.

Editorial extensions

If this is right

  • Users no longer have to write out labeled examples to personalize a recognizer; a few raw handwriting images suffice, removing the annotation bottleneck.
  • Because only the prompts are updated, adaptation can occur on-device or in other memory-constrained settings where full fine-tuning of a 7.6M-parameter model is impractical.
  • The reported 3.36% CER and 10.32% WER on IAM, and 2.19% CER and 6.63% WER on RIMES, are the strongest numbers in the paper's comparison tables, including against methods that personalize with labeled data.
  • Performance improves monotonically as the number of adaptation samples grows from one to five, so the method gets safer with more unlabeled data.
  • Adding prompt layers across the encoder steadily helps, indicating that style information benefits from being injected throughout the feature hierarchy rather than at a single point.

Reading between the lines

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

  • Editorial extension: the same recipe—meta-learned prompt initialization plus a self-supervised inner loop—could be transferred to other tasks where style varies per speaker or writer, such as speech-to-text or music-notation recognition, provided a cheap reconstruction loss is available.
  • The paper evaluates only line-level recognition; page-level documents are a natural next test, but the prompt-padding mechanism would need to encode layout and reading order, which the paper itself notes.
  • The 0.08M parameter count counts only the prompts; a deployment that ships the MAE decoder alongside the frozen recognizer has a larger memory footprint, so the claim of resource efficiency is about adaptation cost, not total model size.
  • A concrete way to stress-test the alignment assumption is to ablate the meta-initialization against a random initialization using the same inner-loop update; if random prompts match the meta-learned ones in CER, the meta-learning step is not the source of the gain.
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 proposes MetaWriter, a method for writer-specific handwritten text recognition that personalizes a pre-trained HTR model by optimizing visual prompt vectors (padding values) using unlabeled test-time images. The prompts are adapted in an inner loop by minimizing a self-supervised SSIM-based masked-autoencoder reconstruction loss, while a meta-learning outer loop trains the initial prompt vectors so that this reconstruction update reduces the text-recognition cross-entropy loss on training writers. The authors report state-of-the-art results on the IAM and RIMES line-level test sets (e.g., 3.36% CER and 10.32% WER on IAM with 5-shot unlabeled adaptation), while updating only 0.08M parameters. The paper includes ablations of the MAE, meta-learning, and prompt components, and a discussion of the limitation to line-level rather than page-level recognition.

Significance. If the reported numbers are reproducible, MetaWriter would be a practical advance: it achieves strong HTR accuracy with a very small number of tunable parameters and removes the need for labeled adaptation data, which is important for deployment. The meta-learned alignment between a self-supervised reconstruction objective and the recognition objective is an interesting and sensible way to engineer a test-time adaptation objective, and the ablation study provides evidence that all three components contribute. The paper is honest about the limitation to line-level recognition. However, the quantitative claims currently rest on an unspecified support-set selection protocol and on inconsistent numbers across tables, and the comparison to non-adaptive baselines is not apples-to-apples; these issues must be resolved before the significance of the results can be assessed.

major comments (4)
  1. [§4.3 (Personalization), Tables 1, 4, 5, 6] The support-set selection rule is unspecified. The text says 'we use a few-shot approach, consistently selecting specific images for each individual rather than relying on random sampling,' but it never defines the selection criterion: whether the five images are the first five lines of each writer, chosen by maximum baseline confidence, a fixed seed, or some other rule. Because every result in the paper depends on this choice, the headline 3.36% CER / 10.32% WER on IAM cannot be independently reproduced or fairly compared against alternative support-set choices. The authors should specify the exact protocol and report means and standard deviations over multiple selections/seeds.
  2. [§5.2, Tables 1, 4, 5, 6] The reported numbers are internally inconsistent. Table 1 reports 3.36% CER / 10.32% WER for the 5-shot IAM setting; Table 4's 'MetaWriter (Ours)' row reports 3.32% / 10.21%; Table 5's L=18 row reports 3.31% / 10.2%; and Table 6 reports 3.36% / 10.32% for K=5 but 3.32% / 10.21% for 'All Samples'. The paper does not state which value is the official result or explain the discrepancies, which directly affects the claimed margin over TrOCR (4.22% CER). The authors must reconcile these values and make the exact evaluation setting unambiguous.
  3. [§5.2, Tables 1 and 2] The comparison to instance-specific training methods (VAN, DAN, TrOCR, etc.) is not apples-to-apples, because those baselines are not given access to test-writer images for adaptation, while MetaWriter receives five unlabeled images from each test writer and updates prompts before inference. The 'outperforming' claim is therefore not established under a matched protocol. The WRA comparison with MetaHTR is also not a direct comparison, since MetaHTR uses word-level labeled adaptation and a different evaluation granularity; the paper should either apply the same adaptation protocol to a strong baseline or clearly restrict the claim to the no-adaptation baseline setting.
  4. [Abstract, §5.2, Table 3] The claim of '20x fewer parameters' is not supported by Table 3. Table 3 reports 0.08M tunable parameters for MetaWriter versus 1.7M for MetaHTR and 7.6M for DAN; the factor is about 21x relative to MetaHTR and about 95x relative to DAN. The '20x' figure appears to assume a specific baseline that is not stated. The paper should state the reference method explicitly and use a consistent definition (e.g., number of tunable parameters during personalization).
minor comments (6)
  1. [§3] Section 3 states 'MetaHTR replies on labeled data'; this should be 'relies on labeled data'.
  2. [§4] Section 4 states 'we propose a efficient framework'; this should be 'an efficient framework'.
  3. [§5.2] Section 5.2 says 'on the world level'; this should be 'on the word level'.
  4. [Figure 4 caption] The figure caption spells 'Origional Image'; it should be 'Original Image'.
  5. [Table 1 and References] Table 1 labels reference [21] as 'TrOCR', but reference [21] is 'Dtrocr: Decoder-only transformer for optical character recognition'; the citation should be corrected or the baseline name changed.
  6. [References] Some references, e.g., [2], [58], and [67], have generic author names that are not standard in the HTR literature and appear unverifiable; the authors should check and replace them.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the meta-training objective is a standard bi-level optimization, and the central claims are validated on held-out test writers rather than derived from the training objective.

full rationale

The paper's central mechanism is a bi-level meta-training procedure (Algorithm 1): the inner loop adapts prompt vectors P_j by one gradient step on the self-supervised reconstruction loss L_ada (Eqs. 1–2), and the outer loop minimizes the supervised recognition loss L_pred on a disjoint query set (Eq. 3; Algorithm 1, lines 5–12). This directly trains the prompt initialization so that a self-supervised update reduces recognition error on training writers; that is a standard meta-learning objective, not a hidden identity or a fitted parameter relabeled as a prediction. The paper's claim about test-time behavior is empirical and is evaluated on held-out IAM and RIMES test writers, which are not used to fit the meta-prompts. The test-time personalization uses only unlabeled images, so the reported CER/WER numbers are not forced by label leakage. The unspecified support-set selection rule in Section 4.3 is a reproducibility and experimental-validity concern, but there is no quoted equation or construction showing the reported numbers reduce to the selection rule, so it does not constitute demonstrated circularity. Self-citations appear in related-work contexts and do not carry the load-bearing argument; no uniqueness theorem from the authors is invoked, and no known result is merely renamed. The paper also includes ablations showing that removing the meta-learning component degrades performance, providing independent support that the mechanism contributes beyond the training objective. Overall, the derivation is self-contained and the held-out evaluations provide genuine external evidence.

Assumptions & free parameters 5 free parameters · 3 assumptions · 0 invented entities

The paper introduces no new particles, forces, dimensions, or other invented entities; the meta-prompts are learnable parameters, not a new entity. The free parameters are standard hyperparameters that the central results depend on, though the authors do not report all fitted values.

free parameters (5)
  • inner loop learning rate lambda1 = not reported; Adam LR = 1e-4 stated
    Controls how far the prompts move during test-time adaptation; central to the method's behavior.
  • outer loop learning rate lambda2 = not reported; Adam LR = 1e-4 stated
    Meta-update step size; not specified separately.
  • support set size k = 5 for main results
    Number of unlabeled images used for adaptation; directly affects error rates (Table 6).
  • number of prompt layers L = 18
    Prompts added to first 18 convolutional layers; Table 5 shows performance varies with L.
  • mask ratio for MAE = 75% (from Figure 4 caption)
    Reconstruction task difficulty depends on mask ratio.
assumptions (3)
  • domain assumption Writer-specific style can be encoded in learnable padding values (prompts) added around input images.
    Section 4.1 assumes style knowledge is captured in prompt vectors; no direct evidence is provided.
  • ad hoc to paper The alignment between the reconstruction loss and recognition loss learned on training writers generalizes to held-out writers.
    Section 4.2 relies on meta-learning for this alignment; its transfer to test writers is the core empirical premise.
  • domain assumption The training set contains enough writers and episodes for the bi-level meta-optimization to find a useful prompt initialization.
    The method requires writer-batched episodic training; if training writers are not representative, adaptation fails.

how reviews work

0 comments
Cite this review

Pith. "Pith review of MetaWriter: Personalized Handwritten Text Recognition Using Meta-Learned Prompt Tuning." pith.science (2026). https://pith.science/paper/RTCUVKJZ

@misc{pith2026250520513,
  author       = {Pith},
  title        = {Pith review of: MetaWriter: Personalized Handwritten Text Recognition Using Meta-Learned Prompt Tuning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/RTCUVKJZ}},
  note         = {Machine review of arXiv:2505.20513}
}
read the original abstract

Recent advancements in handwritten text recognition (HTR) have enabled the effective conversion of handwritten text to digital formats. However, achieving robust recognition across diverse writing styles remains challenging. Traditional HTR methods lack writer-specific personalization at test time due to limitations in model architecture and training strategies. Existing attempts to bridge this gap, through gradient-based meta-learning, still require labeled examples and suffer from parameter-inefficient fine-tuning, leading to substantial computational and memory overhead. To overcome these challenges, we propose an efficient framework that formulates personalization as prompt tuning, incorporating an auxiliary image reconstruction task with a self-supervised loss to guide prompt adaptation with unlabeled test-time examples. To ensure self-supervised loss effectively minimizes text recognition error, we leverage meta-learning to learn the optimal initialization of the prompts. As a result, our method allows the model to efficiently capture unique writing styles by updating less than 1% of its parameters and eliminating the need for time-intensive annotation processes. We validate our approach on the RIMES and IAM Handwriting Database benchmarks, where it consistently outperforms previous state-of-the-art methods while using 20x fewer parameters. We believe this represents a significant advancement in personalized handwritten text recognition, paving the way for more reliable and practical deployment in resource-constrained scenarios.

Figures

Figures reproduced from arXiv: 2505.20513 by the authors.

Figure 1
Figure 1. [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. illustrates the superior performance of our method [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗
Figure 3
Figure 3. Illustration of our approach during training. The handwritten texts from a specific writer are divided into an unlabeled support set and a labeled query set. The images in the support set are masked, padded with meta prompt vectors, and passed through a shared image encoder, followed by reconstruction using a Masked Autoencoder (MAE)’s decoder. The writer-specific prompt vectors Pj are derived in the inner loop usin… view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: Illustration of the capability of Masked AutoEn￾coders to effectively tackle the HTR problem across three di￾verse handwriting styles from test data. The first column repre￾sents the original input images. The second column displays the images masked by 75%, which serv…
Figure 5
Figure 5. Figure 5: illustration of the Character Error Rate (CER) for individual writers (Person IDs 1 to 20) on the IAM dataset. We compare our method with the baseline, showing that our model improves accuracy across all writers [PITH_FULL_IMAGE:figures/full_fig_p008_5.png]
Figure 6
Figure 6. Figure 6: Illustration of the handwriting transcription correc￾tions by MetaWriter from IAM and RIMES. This figure shows original handwriting samples (left), the initial prediction marked with errors highlighted in red (center), and the corrected prediction after applying MetaWr…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

74 extracted references · 59 canonical work pages

  1. [1]

    Improved handwritten digit recognition using convolutional neural networks (cnn)

    Savita Ahlawat, Amit Choudhary, Anand Nayyar, Saurabh Singh, and Byungun Yoon. Improved handwritten digit recognition using convolutional neural networks (cnn). Sen- sors, 20(12):3344, 2020. 2

  2. [2]

    Anderson and Sofia I

    Lucas E. Anderson and Sofia I. Martinez. Applying trans- formers to handwritten text recognition: A new approach. In Proceedings of the International Conference on Machine Learning, pages 785–793, 2022. 2

  3. [3]

    Exploring visual prompts for adapting large- scale models

    Hyojin Bahng, Ali Jahanian, Swami Sankaranarayanan, and Phillip Isola. Exploring visual prompts for adapting large- scale models. arXiv preprint arXiv:2203.17274, 2022. 2

  4. [4]

    Mt3: Meta test-time training for self- supervised test-time adaption

    Alexander Bartler, Andre B ¨uhler, Felix Wiewel, Mario D¨obler, and Bin Yang. Mt3: Meta test-time training for self- supervised test-time adaption. In International Conference on Artificial Intelligence and Statistics , pages 3080–3090. PMLR, 2022. 3

  5. [5]

    Handwriting recognition in low-resource scripts using adversarial learn- ing

    Ayan Kumar Bhunia, Abhirup Das, Ankan Kumar Bhunia, Perla Sai Raj Kishore, and Partha Pratim Roy. Handwriting recognition in low-resource scripts using adversarial learn- ing. In Proceedings of the IEEE/CVF Conference on Com- puter Vision and Pattern Recognition , pages 4767–4776,

  6. [6]

    Metahtr: Towards writer-adaptive handwritten text recogni- tion

    Ayan Kumar Bhunia, Shuvozit Ghose, Amandeep Kumar, Pinaki Nath Chowdhury, Aneeshan Sain, and Yi-Zhe Song. Metahtr: Towards writer-adaptive handwritten text recogni- tion. In Proceedings of the IEEE/CVF Conference on Com- puter Vision and Pattern Recognition , pages 15830–15839,

  7. [7]

    Docsynth: a layout guided approach for controllable docu- ment image synthesis

    Sanket Biswas, Pau Riba, Josep Llad ´os, and Umapada Pal. Docsynth: a layout guided approach for controllable docu- ment image synthesis. In International Conference on Doc- ument Analysis and Recognition , pages 555–568. Springer,

  8. [8]

    Tinytl: Reduce memory, not parameters for efficient on-device learning

    Han Cai, Chuang Gan, Ligeng Zhu, and Song Han. Tinytl: Reduce memory, not parameters for efficient on-device learning. Advances in Neural Information Processing Sys- tems, 33:11285–11297, 2020. 2

Show all 74 references
  1. [9]

    Kartik Chaudhary and Raghav Bali. Easter2. 0: Improving convolutional models for handwritten text recognition.arXiv preprint arXiv:2205.14879, 2022. 6

  2. [10]

    Contrastive test-time adaptation

    Dian Chen, Dequan Wang, Trevor Darrell, and Sayna Ebrahimi. Contrastive test-time adaptation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pat- tern Recognition, pages 295–305, 2022. 3

  3. [11]

    Improved test-time adaptation for domain generalization

    Liang Chen, Yong Zhang, Yibing Song, Ying Shan, and Lingqiao Liu. Improved test-time adaptation for domain generalization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24172– 24182, 2023. 3

  4. [12]

    Test- time fast adaptation for dynamic scene deblurring via meta- auxiliary learning

    Zhixiang Chi, Yang Wang, Yuanhao Yu, and Jin Tang. Test- time fast adaptation for dynamic scene deblurring via meta- auxiliary learning. In Proceedings of the IEEE/CVF Con- ference on Computer Vision and Pattern Recognition, pages 9137–9146, 2021. 3

  5. [13]

    Adapt- ing to distribution shift by visual domain prompt generation

    Zhixiang Chi, Li Gu, Tao Zhong, Huan Liu, YUANHAO YU, Konstantinos N Plataniotis, and Yang Wang. Adapt- ing to distribution shift by visual domain prompt generation. In The Twelfth International Conference on Learning Repre- sentations, 2024. 2

  6. [14]

    Learning to adapt frozen clip for few-shot test-time domain adaptation

    Zhixiang Chi, Li Gu, Huan Liu, Ziqiang Wang, Yanan Wu, Yang Wang, and Konstantinos N Plataniotis. Learning to adapt frozen clip for few-shot test-time domain adaptation. In The Thirteenth International Conference on Learning Representations, 2025. 2

  7. [15]

    Span: a simple predict & align network for handwritten para- graph recognition

    Denis Coquenet, Cl ´ement Chatelain, and Thierry Paquet. Span: a simple predict & align network for handwritten para- graph recognition. In International Conference on Document Analysis and Recognition, pages 70–84. Springer, 2021. 6, 7

  8. [16]

    End-to-end handwritten paragraph text recognition using a vertical attention network

    Denis Coquenet, Cl ´ement Chatelain, and Thierry Paquet. End-to-end handwritten paragraph text recognition using a vertical attention network. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(1):508–524, 2022. 1, 3, 6, 7

  9. [17]

    Dan: a segmentation-free document attention network for handwritten document recognition

    Denis Coquenet, Cl ´ement Chatelain, and Thierry Paquet. Dan: a segmentation-free document attention network for handwritten document recognition. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2023. 3, 6, 7

  10. [18]

    Faster dan: Multi-target queries with document positional encoding for end-to-end handwritten document recognition

    Denis Coquenet, Cl ´ement Chatelain, and Thierry Paquet. Faster dan: Multi-target queries with document positional encoding for end-to-end handwritten document recognition. In International Conference on Document Analysis and Recognition, pages 182–199. Springer, 2023. 1, 3

  11. [19]

    Text and style condi- tioned gan for generation of offline handwriting lines

    Brian Davis, Chris Tensmeyer, Brian Price, Curtis Wiging- ton, Bryan Morse, and Rajiv Jain. Text and style condi- tioned gan for generation of offline handwriting lines. arXiv preprint arXiv:2009.00678, 2020. 1

  12. [20]

    Model- agnostic meta-learning for fast adaptation of deep networks

    Chelsea Finn, Pieter Abbeel, and Sergey Levine. Model- agnostic meta-learning for fast adaptation of deep networks. In International Conference on Machine Learning , pages 1126–1135. PMLR, 2017. 1, 3, 6

  13. [21]

    Dtrocr: Decoder-only transformer for op- tical character recognition

    Masato Fujitake. Dtrocr: Decoder-only transformer for op- tical character recognition. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision , pages 8025–8035, 2024. 1, 6, 7

  14. [22]

    Handwritten gujarati numer- als classification based on deep convolution neural networks using transfer learning scenarios

    Parth Goel and Amit Ganatra. Handwritten gujarati numer- als classification based on deep convolution neural networks using transfer learning scenarios. IEEE Access, 11:20202– 20215, 2023. 6

  15. [23]

    Connectionist temporal classification: labelling unsegmented sequence data with recurrent neural networks

    Alex Graves, Santiago Fern ´andez, Faustino Gomez, and J¨urgen Schmidhuber. Connectionist temporal classification: labelling unsegmented sequence data with recurrent neural networks. In Proceedings of the 23rd International Confer- ence on Machine Learning, pages 369–376, 2006. 2

  16. [24]

    Icdar 2011-french handwriting recognition competition

    Emmanuele Grosicki and Haikal El-Abed. Icdar 2011-french handwriting recognition competition. In 2011 International Conference on Document Analysis and Recognition , pages 1459–1463. IEEE, 2011. 6

  17. [25]

    Improving protonet for few-shot video object recog- nition: Winner of orbit challenge 2022

    Li Gu, Zhixiang Chi, Huan Liu, Yuanhao Yu, and Yang Wang. Improving protonet for few-shot video object recog- nition: Winner of orbit challenge 2022. arXiv preprint arXiv:2210.00174, 2022. 3

  18. [26]

    Masked autoencoders are scalable vision learners

    Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Doll´ar, and Ross Girshick. Masked autoencoders are scalable vision learners. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16000– 16009, 2022. 2, 3

  19. [27]

    Annollm: Making large language mod- els to be better crowdsourced annotators

    Xingwei He, Zhenghao Lin, Yeyun Gong, Alex Jin, Hang Zhang, Chen Lin, Jian Jiao, Siu Ming Yiu, Nan Duan, Weizhu Chen, et al. Annollm: Making large language mod- els to be better crowdsourced annotators. arXiv preprint arXiv:2303.16854, 2023. 2

  20. [28]

    Diversity-aware meta visual prompting

    Qidong Huang, Xiaoyi Dong, Dongdong Chen, Weim- ing Zhang, Feifei Wang, Gang Hua, and Nenghai Yu. Diversity-aware meta visual prompting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10878–10887, 2023. 2

  21. [29]

    Vi- sual prompt tuning

    Menglin Jia, Luming Tang, Bor-Chun Chen, Claire Cardie, Serge Belongie, Bharath Hariharan, and Ser-Nam Lim. Vi- sual prompt tuning. In European Conference on Computer Vision, pages 709–727. Springer, 2022. 2, 3

  22. [30]

    Ganwriting: content- conditioned generation of styled handwritten word images

    Lei Kang, Pau Riba, Yaxing Wang, Marc ¸al Rusinol, Ali- cia Forn ´es, and Mauricio Villegas. Ganwriting: content- conditioned generation of styled handwritten word images. In Computer Vision–ECCV 2020: 16th European Confer- ence, Glasgow, UK, August 23–28, 2020, Proceedings, P...

  23. [31]

    Pay attention to what you read: non- recurrent handwritten text-line recognition

    Lei Kang, Pau Riba, Marc ¸al Rusi ˜nol, Alicia Forn ´es, and Mauricio Villegas. Pay attention to what you read: non- recurrent handwritten text-line recognition. Pattern Recog- nition, 129:108766, 2022. 6

  24. [32]

    Bert: Pre-training of deep bidirectional trans- formers for language understanding

    Jacob Devlin Ming-Wei Chang Kenton and Lee Kristina Toutanova. Bert: Pre-training of deep bidirectional trans- formers for language understanding. In Proceedings of naacL-HLT, page 2. Minneapolis, Minnesota, 2019. 3

  25. [33]

    Recursive recurrent nets with attention modeling for ocr in the wild

    Chen-Yu Lee and Simon Osindero. Recursive recurrent nets with attention modeling for ocr in the wild. In Proceed- ings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 2231–2239, 2016. 2

  26. [34]

    The power of scale for parameter-efficient prompt tuning.arXiv preprint arXiv:2104.08691, 2021

    Brian Lester, Rami Al-Rfou, and Noah Constant. The power of scale for parameter-efficient prompt tuning.arXiv preprint arXiv:2104.08691, 2021. 3

  27. [35]

    Show, attend and read: A simple and strong baseline for irregular text recognition

    Hui Li, Peng Wang, Chunhua Shen, and Guyu Zhang. Show, attend and read: A simple and strong baseline for irregular text recognition. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 8610–8617, 2019. 1

  28. [36]

    Prefix-tuning: Optimiz- ing continuous prompts for generation

    Xiang Lisa Li and Percy Liang. Prefix-tuning: Optimiz- ing continuous prompts for generation. arXiv preprint arXiv:2101.00190, 2021. 3, 4

  29. [37]

    Meta- sgd: Learning to learn quickly for few-shot learning

    Zhenguo Li, Fengwei Zhou, Fei Chen, and Hang Li. Meta- sgd: Learning to learn quickly for few-shot learning. In Pro- ceedings of the 31st International Conference on Neural In- formation Processing Systems, pages 3637–3647, 2017. 3

  30. [38]

    On-device training under 256kb memory

    Ji Lin, Ligeng Zhu, Wei-Ming Chen, Wei-Chen Wang, Chuang Gan, and Song Han. On-device training under 256kb memory. Advances in Neural Information Processing Sys- tems, 35:22941–22954, 2022. 2

  31. [39]

    Scatter: selective con- text attentional scene text recognizer

    Ron Litman, Oron Anschel, Shahar Tsiper, Roee Litman, Shai Mazor, and R Manmatha. Scatter: selective con- text attentional scene text recognizer. In proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11962–11972, 2020. 1

  32. [40]

    Few-shot class-incremental learning via entropy-regularized data-free replay

    Huan Liu, Li Gu, Zhixiang Chi, Yang Wang, Yuanhao Yu, Jun Chen, and Jin Tang. Few-shot class-incremental learning via entropy-regularized data-free replay. In European Con- ference on Computer Vision, pages 146–162. Springer, 2022. 3

  33. [41]

    Meta-auxiliary learning for future depth pre- diction in videos

    Huan Liu, Zhixiang Chi, Yuanhao Yu, Yang Wang, Jun Chen, and Jin Tang. Meta-auxiliary learning for future depth pre- diction in videos. In Proceedings of the IEEE/CVF Win- ter Conference on Applications of Computer Vision , pages 5756–5765, 2023. 3

  34. [42]

    Test-time personalization with meta prompt for gaze estima- tion

    Huan Liu, Julia Qi, Zhenhao Li, Mohammad Hassanpour, Yang Wang, Konstantinos N Plataniotis, and Yuanhao Yu. Test-time personalization with meta prompt for gaze estima- tion. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 3621–3629, 2024. 3, 4

  35. [43]

    Learn to augment: Joint data augmentation and network optimization for text recognition

    Canjie Luo, Yuanzhi Zhu, Lianwen Jin, and Yongpan Wang. Learn to augment: Joint data augmentation and network optimization for text recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13746–13755, 2020. 1, 2

  36. [44]

    Maskocr: Text recognition with masked encoder-decoder pretraining

    Pengyuan Lyu, Chengquan Zhang, Shanshan Liu, Meina Qiao, Yangliu Xu, Liang Wu, Kun Yao, Junyu Han, Er- rui Ding, and Jingdong Wang. Maskocr: Text recognition with masked encoder-decoder pretraining. arXiv preprint arXiv:2206.00311, 2022. 3

  37. [45]

    The iam-database: an english sentence database for offline handwriting recognition

    U-V Marti and Horst Bunke. The iam-database: an english sentence database for offline handwriting recognition. Inter- national journal on document analysis and recognition , 5: 39–46, 2002. 6

  38. [46]

    On first-order meta-learning algorithms

    Alex Nichol, Joshua Achiam, and John Schulman. On first-order meta-learning algorithms. arXiv preprint arXiv:1803.02999, 2018. 3

  39. [47]

    Blackvip: Black-box visual prompting for robust transfer learning

    Changdae Oh, Hyeji Hwang, Hee-young Lee, YongTaek Lim, Geunyoung Jung, Jiyoung Jung, Hosik Choi, and Kyungwoo Song. Blackvip: Black-box visual prompting for robust transfer learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages 2...

  40. [48]

    Meta-learning of pooling layers for character recognition

    Takato Otsuzuki, Heon Song, Seiichi Uchida, and Hideaki Hayashi. Meta-learning of pooling layers for character recognition. In Document Analysis and Recognition–ICDAR 2021: 16th International Conference, Lausanne, Switzer- land, September 5–10, 2021, Proceedings, Part III 16, ...

  41. [49]

    Black box few-shot adaptation for vision- language models

    Yassine Ouali, Adrian Bulat, Brais Matinez, and Georgios Tzimiropoulos. Black box few-shot adaptation for vision- language models. In Proceedings of the IEEE/CVF Interna- tional Conference on Computer Vision, pages 15534–15546,

  42. [50]

    Prompt learning via meta-regularization

    Jinyoung Park, Juyeon Ko, and Hyunwoo J Kim. Prompt learning via meta-regularization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 26940–26950, 2024. 3

  43. [51]

    Joan Puigcerver. Are multidimensional recurrent layers re- ally necessary for handwritten text recognition? In 2017 14th IAPR International Conference on Document Analysis and Recognition (ICDAR), pages 67–72. IEEE, 2017. 6

  44. [52]

    Meta self- learning for multi-source domain adaptation: a benchmark

    Shuhao Qiu, Chuang Zhu, and Wenli Zhou. Meta self- learning for multi-source domain adaptation: a benchmark. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1592–1601, 2021. 3

  45. [53]

    Learn- ing transferable visual models from natural language super- vision

    Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learn- ing transferable visual models from natural language super- vision. In International Conference on Machine Learning...

  46. [54]

    Transformer-based approach for joint handwriting and named entity recognition in historical doc- ument

    Ahmed Cheikh Rouhou, Marwa Dhiaf, Yousri Kessentini, and Sinda Ben Salem. Transformer-based approach for joint handwriting and named entity recognition in historical doc- ument. Pattern Recognition Letters, 155:128–134, 2022. 6

  47. [55]

    Image quality assessment through fsim, ssim, mse and psnr—a comparative study

    Umme Sara, Morium Akter, and Mohammad Shorif Ud- din. Image quality assessment through fsim, ssim, mse and psnr—a comparative study. Journal of Computer and Com- munications, 7(3):8–18, 2019. 4

  48. [56]

    Psnr vs ssim: impercepti- bility quality assessment for image steganography

    De Rosal Igantius Moses Setiadi. Psnr vs ssim: impercepti- bility quality assessment for image steganography. Multime- dia Tools and Applications, 80(6):8423–8444, 2021. 4

  49. [57]

    Aster: An attentional scene text recognizer with flexible rectification

    Baoguang Shi, Mingkun Yang, Xinggang Wang, Pengyuan Lyu, Cong Yao, and Xiang Bai. Aster: An attentional scene text recognizer with flexible rectification. IEEE transactions on pattern analysis and machine intelligence , 41(9):2035– 2048, 2018. 2

  50. [58]

    Smith and Jane B

    John A. Smith and Jane B. Doe. Enhanced feature extrac- tion in handwritten document recognition using deep convo- lutional neural networks. Journal of Computer Vision and Image Understanding, 150:102–114, 2021. 2

  51. [59]

    Test-time training with self- supervision for generalization under distribution shifts

    Yu Sun, Xiaolong Wang, Zhuang Liu, John Miller, Alexei Efros, and Moritz Hardt. Test-time training with self- supervision for generalization under distribution shifts. InIn- ternational Conference on Machine Learning , pages 9229–

  52. [60]

    Attention is all you need

    Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszko- reit, Llion Jones, Aidan N Gomez, Łukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in Neural Information Processing Systems, 30, 2017. 3

  53. [61]

    Tent: Fully test-time adaptation by entropy minimization

    Dequan Wang, Evan Shelhamer, Shaoteng Liu, Bruno Ol- shausen, and Trevor Darrell. Tent: Fully test-time adaptation by entropy minimization. arXiv preprint arXiv:2006.10726,

  54. [62]

    Decoupled attention network for text recognition

    Tianwei Wang, Yuanzhi Zhu, Lianwen Jin, Canjie Luo, Xi- aoxue Chen, Yaqiang Wu, Qianying Wang, and Mingxiang Cai. Decoupled attention network for text recognition. In Proceedings of the AAAI Conference on Artificial Intelli- gence, pages 12216–12224, 2020. 2, 6

  55. [63]

    Learning to prompt for continual learning

    Zifeng Wang, Zizhao Zhang, Chen-Yu Lee, Han Zhang, Ruoxi Sun, Xiaoqi Ren, Guolong Su, Vincent Perot, Jennifer Dy, and Tomas Pfister. Learning to prompt for continual learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages 139–149,

  56. [64]

    Distribution align- ment for fully test-time adaptation with dynamic online data streams

    Ziqiang Wang, Zhixiang Chi, Yanan Wu, Li Gu, Zhi Liu, Konstantinos Plataniotis, and Yang Wang. Distribution align- ment for fully test-time adaptation with dynamic online data streams. In European Conference on Computer Vision, pages 332–349. Springer, 2024. 3

  57. [65]

    Fast writer adaptation with style extractor network for handwritten text recognition

    Zi-Rui Wang and Jun Du. Fast writer adaptation with style extractor network for handwritten text recognition. Neural Networks, 147:42–52, 2022. 1, 6

  58. [66]

    Start, follow, read: End-to-end full-page handwriting recognition

    Curtis Wigington, Chris Tensmeyer, Brian Davis, William Barrett, Brian Price, and Scott Cohen. Start, follow, read: End-to-end full-page handwriting recognition. In Pro- ceedings of the European Conference on Computer Vision (ECCV), pages 367–383, 2018. 6

  59. [67]

    Williams and Neha S

    David L. Williams and Neha S. Patel. A comprehensive re- view of large-scale datasets for offline handwritten document recognition. IEEE Transactions on Pattern Analysis and Ma- chine Intelligence, 45(2):567–579, 2023. 2

  60. [68]

    Metagcd: Learning to continually learn in generalized cat- egory discovery

    Yanan Wu, Zhixiang Chi, Yang Wang, and Songhe Feng. Metagcd: Learning to continually learn in generalized cat- egory discovery. In Proceedings of the IEEE/CVF Interna- tional Conference on Computer Vision , pages 1655–1665,

  61. [69]

    Test-time domain adaptation by learning domain-aware batch normalization

    Yanan Wu, Zhixiang Chi, Yang Wang, Konstantinos N Pla- taniotis, and Songhe Feng. Test-time domain adaptation by learning domain-aware batch normalization. In Proceed- ings of the AAAI Conference on Artificial Intelligence, pages 15961–15969, 2024. 3

  62. [70]

    A-vit: Adaptive tokens for efficient vision transformer

    Hongxu Yin, Arash Vahdat, Jose M Alvarez, Arun Mallya, Jan Kautz, and Pavlo Molchanov. A-vit: Adaptive tokens for efficient vision transformer. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages 10809–10818, 2022. 3

  63. [71]

    Robust test- time adaptation in dynamic scenarios

    Longhui Yuan, Binhui Xie, and Shuang Li. Robust test- time adaptation in dynamic scenarios. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15922–15932, 2023. 3

  64. [72]

    Sequence-to-sequence domain adaptation network for robust text image recognition

    Yaping Zhang, Shuai Nie, Wenju Liu, Xing Xu, Dongxiang Zhang, and Heng Tao Shen. Sequence-to-sequence domain adaptation network for robust text image recognition. InPro- ceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2740–2749, 2019. 2

  65. [73]

    Meta-dmoe: Adapting to domain shift by meta- distillation from mixture-of-experts

    Tao Zhong, Zhixiang Chi, Li Gu, Yang Wang, Yuanhao Yu, and Jin Tang. Meta-dmoe: Adapting to domain shift by meta- distillation from mixture-of-experts. Advances in Neural In- formation Processing Systems, 35:22243–22257, 2022. 3

  66. [74]

    Learning to prompt for vision-language models

    Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Learning to prompt for vision-language models. In- ternational Journal of Computer Vision, 130(9):2337–2348,

Pith tools

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