Pith. sign in

REVIEW 5 major objections 6 minor 38 references

Col-OLHTR: A Novel Framework for Multimodal Online Handwritten Text Recognition

T0 review · 5 major / 6 minor · reviewed 2026-08-08 · deepseek-v4-flash

Pith's one-line read Col-OLHTR trains a trajectory-only handwriting recognizer with an image teacher, then deletes the teacher at inference, reporting lower error than multi-stream fusion.

desk verdict Plausible engineering result with consistent gains, but the P2SA mechanism is not supported as written: the image teacher is height-collapsed, so the alignment loss can only teach column features, not 2D image structure. read the letter →

arxiv 2502.06100 v1 pith:SRN4YHSQ submitted 2025-02-10 cs.CV eess.SP

classification cs.CVeess.SP
keywords onlinehandwrittentextrecognitiontrajectorycollaborativelearningpoint-to-spatialalignment2Drotarypositionembeddingmultimodalfusionsingle-streaminferenceattentiondecoder
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

Col-OLHTR is an online handwritten text recognition framework built around a deliberate swap: use an image encoder during training to teach the trajectory encoder spatial awareness, then discard the image stream at inference. The paper argues that a single trajectory stream plus a Point-to-Spatial Alignment (P2SA) module can match or beat multi-stream fusion models, reporting 4.34% character error rate on IAM-OnDB against 4.70% for the best compared fusion method, and 95.02% accuracy on ICDAR2013-Online against 93.92%. If the claim holds, the accuracy of bimodal fusion can be obtained without paying the inference cost of rendering images and running a second encoder.

What carries the argument

The Point-to-Spatial Alignment (P2SA) module is the load-bearing component. It takes the downsampled 1D convolutional features $F_{\mathrm{conv}}^{1d}$, adds a 2D rotary position embedding built from the raw pen coordinates $S$ to obtain $F_{\mathrm{pos}}^{1d}$, and passes this through three Transformer layers to produce $F_{p2s}$. The training target is $F_{\mathrm{sample}}^{2d} = \mathrm{Interpolation}(F_{\mathrm{conv}}^{2d}, S)$, the image encoder's features linearly interpolated at the trajectory point positions, and the loss $\mathcal{L}_{\mathrm{align}} = \|F_{p2s} - \mathrm{SG}(F_{\mathrm{sample}}^{2d})\|_2^2$ pushes P2SA to reproduce image-level spatial structure from point positions alone. The stop-gradient operation prevents the image stream from being dragged toward the trajectory stream, and the final trajectory features are formed by adding $F_{p2s}$ back into the 1D encoder before the BiGRU layers.

What would settle it

Retrain the full model on IAM-OnDB with the alignment loss removed but the P2SA Transformer and RoPE kept; if the CER stays near 4.34% rather than climbing toward the 5.00% ablation row, the alignment loss is not the cause of the reported gain. A second check is to vary the fixed 32-pixel rendering height down to, say, 16 pixels; if the gap over a trajectory-only baseline disappears, the result depends on rendering resolution rather than on the alignment idea itself.

Watch

Extended reading notes

Core claim

The central claim is that multimodal features can be learned during training and collapsed into a single-stream inference model. The P2SA module maps trajectory-level features into image-level features using 2D rotary position embeddings and three Transformer layers; during training, an MSE loss aligns these mapped features with point-sampled features from a concurrently trained image encoder, and a stop-gradient keeps the image stream fixed as a teacher. At test time only the trajectory encoder, P2SA, and attention decoder remain, so the pipeline is simpler than multi-stream fusion while its reported accuracy is higher on the benchmarks considered, including a 4.34% CER on IAM-OnDB where the best compared multi-stream method gives 4.70%.

Load-bearing premise

The image features sampled at trajectory point positions carry enough spatial structure to teach P2SA useful image-level features; if that interpolation loses stroke-shape information, the collaborative supervision cannot give the trajectory stream the spatial awareness the paper claims.

Editorial extensions

If this is right

  • With only raw trajectory input at inference, Col-OLHTR reports a lower character error rate than the compared multi-stream fusion method on IAM-OnDB (4.34% vs 4.70%), so the image stream can be removed without giving up accuracy.
  • On the short-word OnHW-WordsTraj benchmark it reports the best CER and WER among the compared methods (1.61% and 11.24%).
  • On the sloppy-style Chinese ICDAR2013-Online set it reports 95.02% AR and 95.34% CR, above every compared fusion method.
  • At inference the forward pass reduces to Eqn. (1), Eqn. (5), Eqn. (8), and Eqn. (9), so no image rendering or second encoder-decoder is required.
  • The ablation attributes the largest single improvement to the stop-gradient alignment loss: adding it after the Transformer and RoPE lowers CER from 4.92% to 4.34%.

Reading between the lines

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

  • The same collaborative-supervision recipe could transfer to other sequence-to-structure tasks with known coordinate tracks and renderable surrogates, such as mathematical expression or sketch recognition, giving those systems spatial awareness at training time without a second stream at inference.
  • Because the method is effectively feature distillation through an MSE loss with a stop-gradient teacher, it suggests a generic alternative to classifier-based distillation that needs no extra task head during training.
  • The fixed 32-pixel rendering height and the choice to sample at trajectory points are untested hyperparameters; a natural extension is to ablate rendering resolution and point density to see how much spatial detail the alignment actually preserves.
  • The single MSE alignment leaves open whether a richer multimodal attention interaction during training could give the trajectory stream an even larger spatial vocabulary at zero extra inference cost.
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 / 6 minor

Summary. Col-OLHTR is a collaborative-learning framework for online handwritten text recognition (OLHTR). During training, a trajectory stream (1D CNN + BiGRU encoder and attention decoder) is supplemented by a Point-to-Spatial Alignment (P2SA) module and an auxiliary image stream (ResNet-18 2D CNN + BiGRU encoder and attention decoder). P2SA adds 2D rotary position embeddings to the temporally downsampled trajectory features and passes them through three Transformer layers; an MSE alignment loss encourages P2SA features to match image-encoder features sampled at the trajectory point locations, with a stop-gradient on the target. At inference the image stream is discarded, and recognition uses only the trajectory stream and P2SA. Experiments report CER/WER of 4.34%/17.23% on IAM-OnDB, 1.61%/11.24% on OnHW-WordsTraj, and AR/CR of 95.02%/95.34% on ICDAR2013-Online, outperforming the cited multi-stream baseline Multi-scale Bi-Fusion. An ablation on IAM-OnDB decomposes the gains of the Transformer, RoPE, alignment loss, and stop-gradient.

Significance. If the results are reproducible, the paper's practical contribution is clear: a single-stream trajectory-only inference model that matches or beats multi-stream fusion baselines (e.g., CER 4.34 vs 4.70 on IAM-OnDB and AR 95.02 vs 93.92 on ICDAR2013-Online) while avoiding image rendering and extra streams at test time. The framework is simple, the ablation on IAM-OnDB supports the usefulness of the P2SA components when applied cumulatively, and the comparison spans English and Chinese benchmarks. At the same time, the central mechanistic claim — that P2SA learns genuine 2D image-level spatial features — is not established by the stated architecture because the 2D encoder's feature map is height-collapsed. The SOTA claim is also weakened by single-run results without uncertainty estimates and by an incomplete description of the synthetic data used for the Chinese line-recognition experiment.

major comments (5)
  1. [Section II-B / II-C, Eqns. (3), (6)-(7)] Claimed 2D spatial supervision is not supported by the stated architecture. With input image I in R^{32xW} and a ResNet-18 of total stride (32,8), Eqn. (3) produces Fconv_2d in R^{W/8 x d}, i.e., the height dimension is a single row. Consequently, the linear interpolation in Eqn. (6) samples features with no vertical index, and the MSE target in Eqn. (7) cannot teach P2SA any y-specific image content. The 2D RoPE on the student side supplies y-coordinates, but those are derived from the trajectory itself, not from the image teacher. The authors should either change the 2D encoder to preserve a non-trivial height dimension, or provide direct evidence (e.g., feature visualizations or a height-preserving teacher ablation) that the P2SA output learns spatial structure beyond the raw (x,y) point coordinates.
  2. [Section II-C, Eqns. (5)-(8)] Shape mismatches around the P2SA module are unspecified. Fconv_1d in Eqn. (1) is in R^{T/8 x d}, while RoPE(S) in Eqn. (5) operates on the raw trajectory S in R^{T x 3}; the paper does not state how the positions are downsampled or interpolated to the T/8 time steps before the addition. Likewise, Eqn. (6) samples from Fconv_2d in R^{W/8 x d} at "points in S", but the relationship between T and W and the exact sampling coordinates are not given, and Fp2s must be length T/8 to be added in Eqn. (8). These hidden resampling choices may affect the reported gains and should be stated explicitly and ideally ablated.
  3. [Section III-D, Table IV] Table IV does not include the isolated alignment-loss ablation described in the text. The sentence "using only the additional Lalign did not yield improvements" has no supporting row: the closest configuration is Transformer+RoPE+AlignLoss (without SG) at 4.92, which is better than Transformer+RoPE at 5.00. If the intended comparison is row 4 versus row 3, the improvement is small but still positive; if the intended comparison is baseline plus Lalign alone, that row is missing. Please add the missing configuration or revise the wording.
  4. [Section III-B / III-C] All experiments are reported from single runs without error bars, confidence intervals, or significance tests. Since the SOTA claims rest on differences such as CER 4.34 vs 4.70 and AR 95.02 vs 93.92, the authors should report the mean and standard deviation over at least three random seeds for the main results and ablations, or provide a statistical test, to establish that the improvements are not due to training variance.
  5. [Section III-A / III-C] The synthetic training data used for the ICDAR2013-Online experiment is not described. Section III-C states that the model is trained on CASIA-OLHWDB2.0-2.2 plus "synthesized samples generated from the single characters of CASIA-OLHWDB1.0-1.2", but no details of the synthesis (line construction, rendering, resizing, augmentation, or amount) are provided. This is necessary to reproduce the reported AR/CR of 95.02%/95.34% and to ensure a fair comparison with Multi-scale Bi-Fusion. Please give the full protocol or a reference to it.
minor comments (6)
  1. [Section II-C, Eqn. (7)] The stop-gradient is applied to the target Fsample_2d, but Section III-D refers to the "stop-gradient design on the 2D feature maps"; the wording should be aligned with the equation.
  2. [Section II-C] The term "RoiAlign" should be spelled "RoIAlign".
  3. [Section II-B, Eqn. (3)] The use of a ResNet-18 with total stride (32,8) is unconventional; please specify the stem and max-pooling configuration or cite an implementation.
  4. [Section III-B] The data augmentation is described as "random perturbations applied to 20% of stroke points" without specifying the perturbation distribution; please provide this detail for reproducibility.
  5. [Table IV] The table column headers use "Transformer" and "RoPE" without the number of layers (three) given in Section II-C; adding the layer count in the caption would make the table self-contained.
  6. [Section III-D] The ablation is performed only on IAM-OnDB; adding at least one Chinese dataset row would strengthen the generality claim.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the SOTA claim is supported by held-out benchmark evaluation, and the auxiliary alignment loss is training-only supervision rather than a re-reporting of fitted values.

full rationale

This is an empirical systems paper with no derivation whose output is defined by its input. The central claim is evaluated on held-out test sets (IAM-OnDB, OnHW-WordsTraj, ICDAR2013-Online) against external published methods using standard CER/WER/AR metrics. The P2SA alignment target in Eqns. (6)-(7) is an internal training-time MSE supervision obtained from the image encoder, not a reported prediction; the stop-gradient operation in Eqn. (7) explicitly prevents the image encoder from being directly fitted to the student features, and at inference the image stream is discarded. No fitted constant is presented as a prediction: the hyperparameter lambda is set to 2.0, and design choices are supported by ablations (Table IV). The few references to the authors' prior work are standard baselines or background material and are not load-bearing for the SOTA claim. The skeptics' concern about the height-collapsing ResNet-18 stride is a legitimate correctness/mechanism question, but it is not circularity: the reported numbers may be correct while the claimed mechanism is under-specified. No self-citation chain, uniqueness import, or ansatz-by-citation appears in the derivation. Therefore the appropriate circularity score is 0.

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

The paper contributes no new physical entities; P2SA is an architectural module whose evidence is the ablation. The main hand-set quantities are the alignment loss weight, P2SA depth, feature dimension, and rendering height. The central assumption is that point-sampled image features are a valid spatial target for trajectory features, and that RoPE added to point features supplies the missing 2D structure. These are reasonable engineering assumptions but they are not independently verified outside the reported experiments.

free parameters (5)
  • lambda_align = 2.0
    Weight of the P2SA alignment loss in Eqn. (11), chosen by hand with no sensitivity study; the central training objective depends on it.
  • p2sa_transformer_layers = 3
    Number of Transformer encoder layers in the P2SA module, fixed by design; ablation shows adding them alone gives only a small CER drop.
  • feature_dim = 320
    Encoder feature dimension d in Section II-B1, a hand-set capacity choice that affects all comparisons.
  • rendered_image_height = 32 pixels
    All training images are resized to height 32, which defines the spatial resolution the P2SA target comes from.
  • data_augmentation_ratio = 20% of stroke points
    Random perturbation ratio stated in Section III-B; not swept in the paper.
assumptions (5)
  • standard math Standard deep-learning building blocks (1D/2D convolutions, BiGRU, Transformer attention, GRU decoder) behave as described in their source papers.
    Section II-B and II-D rely on these components without re-deriving them.
  • domain assumption Rendering the online trajectory to an image and sampling image features at trajectory points via linear interpolation preserves the spatial information needed for recognition.
    Section II-C, Eqn. (6), assumes each trajectory point lies in the rendered image and that point-level interpolation of Fconv_2d is a good alignment target.
  • domain assumption The auxiliary image stream's features are a useful supervision source for the trajectory stream, and stop-gradient prevents the image stream from being corrupted.
    The collaborative loss in Eqn. (7) is the mechanism that transfers spatial knowledge; the paper does not analyze failure cases.
  • domain assumption Public benchmark splits and evaluation protocols for IAM-OnDB, OnHW-WordsTraj, CASIA-OLHWDB, and ICDAR2013 are followed correctly.
    All conclusions depend on standard splits and metrics, and no preprocessing code is released to verify them.
  • ad hoc to paper Adding 2D rotary position embeddings directly to point features, as in Eqn. (5), gives the model spatial awareness without additional image input.
    This design choice is specific to Col-OLHTR and is justified only by the ablation result, not by a theoretical argument.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Col-OLHTR: A Novel Framework for Multimodal Online Handwritten Text Recognition." pith.science (2026). https://pith.science/paper/SRN4YHSQ

@misc{pith2026250206100,
  author       = {Pith},
  title        = {Pith review of: Col-OLHTR: A Novel Framework for Multimodal Online Handwritten Text Recognition},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/SRN4YHSQ}},
  note         = {Machine review of arXiv:2502.06100}
}
read the original abstract

Online Handwritten Text Recognition (OLHTR) has gained considerable attention for its diverse range of applications. Current approaches usually treat OLHTR as a sequence recognition task, employing either a single trajectory or image encoder, or multi-stream encoders, combined with a CTC or attention-based recognition decoder. However, these approaches face several drawbacks: 1) single encoders typically focus on either local trajectories or visual regions, lacking the ability to dynamically capture relevant global features in challenging cases; 2) multi-stream encoders, while more comprehensive, suffer from complex structures and increased inference costs. To tackle this, we propose a Collaborative learning-based OLHTR framework, called Col-OLHTR, that learns multimodal features during training while maintaining a single-stream inference process. Col-OLHTR consists of a trajectory encoder, a Point-to-Spatial Alignment (P2SA) module, and an attention-based decoder. The P2SA module is designed to learn image-level spatial features through trajectory-encoded features and 2D rotary position embeddings. During training, an additional image-stream encoder-decoder is collaboratively trained to provide supervision for P2SA features. At inference, the extra streams are discarded, and only the P2SA module is used and merged before the decoder, simplifying the process while preserving high performance. Extensive experimental results on several OLHTR benchmarks demonstrate the state-of-the-art (SOTA) performance, proving the effectiveness and robustness of our design.

Figures

Figures reproduced from arXiv: 2502.06100 by the authors.

Figure 1
Figure 1. Overall training framework of the proposed Col-OLHTR. During testing, only the bottom trajectory stream and the P2SA module are utilized. [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

38 extracted references · 28 canonical work pages

  1. [1]

    Imagenet classification with deep convolutional neural networks,

    A. Krizhevsky, I. Sutskever, and G. E. Hinton, “Imagenet classification with deep convolutional neural networks,” Advances in neural informa- tion processing systems , vol. 25, 2012

  2. [2]

    Deep residual learning for image recognition,

    K. He, X. Zhang, S. Ren, and J. Sun, “Deep residual learning for image recognition,” in Proceedings of the IEEE conference on computer vision and pattern recognition , 2016, pp. 770–778

  3. [3]

    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,” Advances in neural information processing systems , vol. 30, 2017

  4. [4]

    Connection- ist temporal classification: labelling unsegmented sequence data with recurrent neural networks,

    A. Graves, S. Fetrn ´andez, F. Gomez, and J. Schmidhuber, “Connection- ist temporal classification: labelling unsegmented sequence data with recurrent neural networks,” in Proceedings of the 23rd international conference on Machine learning , 2006, pp. 369–376

  5. [5]

    Neural machine translation by jointly learning to align and translate,

    D. Bahdanau, K. Cho, and Y . Bengio, “Neural machine translation by jointly learning to align and translate,” arXiv preprint arXiv:1409.0473 , 2014

  6. [6]

    Listen, attend and spell: A neural network for large vocabulary conversational speech recognition,

    W. Chan, N. Jaitly, Q. Le, and O. Vinyals, “Listen, attend and spell: A neural network for large vocabulary conversational speech recognition,” in 2016 IEEE international conference on acoustics, speech and signal processing (ICASSP). IEEE, 2016, pp. 4960–4964

  7. [7]

    Modeling coverage for neural machine translation,

    Z. Tu, Z. Lu, Y . Liu, X. Liu, and H. Li, “Modeling coverage for neural machine translation,” arXiv preprint arXiv:1601.04811 , 2016

  8. [8]

    Deep blstm neural networks for uncon- strained continuous handwritten text recognition,

    V . Frinken and S. Uchida, “Deep blstm neural networks for uncon- strained continuous handwritten text recognition,” in 2015 13th inter- national conference on document analysis and recognition (ICDAR) . IEEE, 2015, pp. 911–915

Show all 38 references
  1. [9]

    Deep neural network based hidden markov model for offline handwritten chinese text recognition,

    J. Du, Z.-R. Wang, J.-F. Zhai, and J.-S. Hu, “Deep neural network based hidden markov model for offline handwritten chinese text recognition,” in 2016 23rd International Conference on Pattern Recognition (ICPR) . IEEE, 2016, pp. 3428–3433

  2. [10]

    Multi-language online handwriting recognition,

    D. Keysers, T. Deselaers, H. A. Rowley, L.-L. Wang, and V . Carbune, “Multi-language online handwriting recognition,” IEEE transactions on pattern analysis and machine intelligence , vol. 39, no. 6, pp. 1180–1194, 2016

  3. [11]

    Learning spatial-semantic context with fully convolutional recurrent network for online handwritten chinese text recognition,

    Z. Xie, Z. Sun, L. Jin, H. Ni, and T. Lyons, “Learning spatial-semantic context with fully convolutional recurrent network for online handwritten chinese text recognition,” IEEE transactions on pattern analysis and machine intelligence, vol. 40, no. 8, pp. 1903–1917, 2017

  4. [12]

    Writer adaptation via deeply learned features for online chinese handwriting recognition,

    J. Du, J.-F. Zhai, and J.-S. Hu, “Writer adaptation via deeply learned features for online chinese handwriting recognition,” International Jour- nal on Document Analysis and Recognition (IJDAR) , vol. 20, pp. 69–78, 2017

  5. [13]

    Fast multi-language lstm-based online handwriting recognition,

    V . Carbune, P. Gonnet, T. Deselaers, H. A. Rowley, A. Daryin, M. Calvo, L.-L. Wang, D. Keysers, S. Feuz, and P. Gervais, “Fast multi-language lstm-based online handwriting recognition,” International Journal on Document Analysis and Recognition (IJDAR) , vol. 23, no. 2, pp. 8...

  6. [14]

    Inceptiontime: Finding alexnet for time series classification,

    H. Ismail Fawaz, B. Lucas, G. Forestier, C. Pelletier, D. F. Schmidt, J. Weber, G. I. Webb, L. Idoumghar, P.-A. Muller, and F. Petitjean, “Inceptiontime: Finding alexnet for time series classification,” Data Mining and Knowledge Discovery , vol. 34, no. 6, pp. 1936–1962, 2020

  7. [15]

    Towards fast, accurate and compact online handwritten chinese text recognition,

    D. Peng, C. Xie, H. Li, L. Jin, Z. Xie, K. Ding, Y . Huang, and Y . Wu, “Towards fast, accurate and compact online handwritten chinese text recognition,” in Document Analysis and Recognition–ICDAR 2021: 16th International Conference, Lausanne, Switzerland, September 5–10, 2021...

  8. [16]

    A multi- scale bimodal fusion network for robust and accurate online handwriting recognition,

    Z. Xu, Z. Chen, Y . Wu, H. Li, W. Lv, L. Jin, and Q. Wang, “A multi- scale bimodal fusion network for robust and accurate online handwriting recognition,” in ICASSP 2024-2024 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP) . IEEE, 2024, pp. 6460–6464

  9. [17]

    A novel connectionist system for unconstrained hand- writing recognition,

    A. Graves, M. Liwicki, S. Fern ´andez, R. Bertolami, H. Bunke, and J. Schmidhuber, “A novel connectionist system for unconstrained hand- writing recognition,” IEEE transactions on pattern analysis and machine intelligence, vol. 31, no. 5, pp. 855–868, 2008

  10. [18]

    In-air handwritten chinese text recognition with temporal convolutional recurrent network,

    J. Gan, W. Wang, and K. Lu, “In-air handwritten chinese text recognition with temporal convolutional recurrent network,” Pattern Recognition , vol. 97, p. 107025, 2020

  11. [19]

    Writer-aware cnn for parsimo- nious hmm-based offline handwritten chinese text recognition,

    Z.-R. Wang, J. Du, and J.-M. Wang, “Writer-aware cnn for parsimo- nious hmm-based offline handwritten chinese text recognition,” Pattern Recognition, vol. 100, p. 107102, 2020

  12. [20]

    Text-conditioned character segmentation for ctc-based text recognition,

    R. Tanaka, K. Osada, and A. Furuhata, “Text-conditioned character segmentation for ctc-based text recognition,” in Document Analysis and Recognition–ICDAR 2021: 16th International Conference, Lausanne, Switzerland, September 5–10, 2021, Proceedings, Part III 16 . Springer, 202...

  13. [21]

    Recognition of handwritten chinese text by segmentation: a segment- annotation-free approach,

    D. Peng, L. Jin, W. Ma, C. Xie, H. Zhang, S. Zhu, and J. Li, “Recognition of handwritten chinese text by segmentation: a segment- annotation-free approach,” IEEE Transactions on Multimedia , vol. 25, pp. 2368–2381, 2022

  14. [22]

    A study on the use of 8-directional features for online handwritten chinese character recognition,

    Z.-L. Bai and Q. Huo, “A study on the use of 8-directional features for online handwritten chinese character recognition,” in Eighth Interna- tional Conference on Document Analysis and Recognition (ICDAR’05) . IEEE, 2005, pp. 262–266

  15. [23]

    Handwritten text recognition with convolutional prototype network and most aligned frame based ctc training,

    L. Gao, H. Zhang, and C.-L. Liu, “Handwritten text recognition with convolutional prototype network and most aligned frame based ctc training,” in Document Analysis and Recognition–ICDAR 2021: 16th In- ternational Conference, Lausanne, Switzerland, September 5–10, 2021, Procee...

  16. [24]

    A compact cnn-dblstm based character model for online handwritten chinese text recognition,

    K. Chen, L. Tian, H. Ding, M. Cai, L. Sun, S. Liang, and Q. Huo, “A compact cnn-dblstm based character model for online handwritten chinese text recognition,” in 2017 14th IAPR international conference on document analysis and Recognition (ICDAR) , vol. 1. IEEE, 2017, pp. 1068–1073

  17. [25]

    Integration of paths–a faithful representation of paths by noncommutative formal power series,

    K.-T. Chen, “Integration of paths–a faithful representation of paths by noncommutative formal power series,” Transactions of the American Mathematical Society , vol. 89, no. 2, pp. 395–407, 1958

  18. [26]

    Indic handwritten script identification using offline-online multi-modal deep network,

    A. K. Bhunia, S. Mukherjee, A. Sain, A. K. Bhunia, P. P. Roy, and U. Pal, “Indic handwritten script identification using offline-online multi-modal deep network,” Information Fusion , vol. 57, pp. 1–14, 2020

  19. [27]

    Multi-modal attention network for handwritten mathematical expression recognition,

    J. Wang, J. Du, J. Zhang, and Z.-R. Wang, “Multi-modal attention network for handwritten mathematical expression recognition,” in 2019 International Conference on Document Analysis and Recognition (IC- DAR). IEEE, 2019, pp. 1181–1186

  20. [28]

    Empirical evaluation of gated recurrent neural networks on sequence modeling,

    J. Chung, C. Gulcehre, K. Cho, and Y . Bengio, “Empirical evaluation of gated recurrent neural networks on sequence modeling,” arXiv preprint arXiv:1412.3555, 2014

  21. [29]

    Roformer: En- hanced transformer with rotary position embedding,

    J. Su, M. Ahmed, Y . Lu, S. Pan, W. Bo, and Y . Liu, “Roformer: En- hanced transformer with rotary position embedding,” Neurocomputing, vol. 568, p. 127063, 2024

  22. [30]

    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 , 2017, pp. 2961–2969

  23. [31]

    Iam-ondb-an on-line english sentence database acquired from handwritten text on a whiteboard,

    M. Liwicki and H. Bunke, “Iam-ondb-an on-line english sentence database acquired from handwritten text on a whiteboard,” in Eighth International Conference on Document Analysis and Recognition (IC- DAR’05). IEEE, 2005, pp. 956–961

  24. [32]

    The onhw dataset: Online handwriting recognition from imu-enhanced ballpoint pens with machine learning,

    F. Ott, M. Wehbi, T. Hamann, J. Barth, B. Eskofier, and C. Mutschler, “The onhw dataset: Online handwriting recognition from imu-enhanced ballpoint pens with machine learning,” Proceedings of the ACM on Interactive, Mobile, Wearable and Ubiquitous Technologies, vol. 4, no. 3, ...

  25. [33]

    Casia online and offline chinese handwriting databases,

    C.-L. Liu, F. Yin, D.-H. Wang, and Q.-F. Wang, “Casia online and offline chinese handwriting databases,” in 2011 international conference on document analysis and recognition . IEEE, 2011, pp. 37–41

  26. [34]

    Icdar 2013 chinese handwriting recognition competition,

    F. Yin, Q.-F. Wang, X.-Y . Zhang, and C.-L. Liu, “Icdar 2013 chinese handwriting recognition competition,” in 2013 12th international con- ference on document analysis and recognition . IEEE, 2013, pp. 1464– 1470

  27. [35]

    Adam: A method for stochastic optimization,

    D. P. Kingma and J. Ba, “Adam: A method for stochastic optimization,” arXiv preprint arXiv:1412.6980 , 2014

  28. [36]

    Benchmarking online sequence-to-sequence and character-based handwriting recognition from imu-enhanced pens,

    F. Ott, D. R ¨ugamer, L. Heublein, T. Hamann, J. Barth, B. Bischl, and C. Mutschler, “Benchmarking online sequence-to-sequence and character-based handwriting recognition from imu-enhanced pens,” In- ternational Journal on Document Analysis and Recognition (IJDAR) , vol. 25, n...

  29. [37]

    Combining diverse systems for handwritten text line recognition,

    M. Liwicki, H. Bunke, J. A. Pittman, and S. Knerr, “Combining diverse systems for handwritten text line recognition,” Machine vision and applications, vol. 22, pp. 39–51, 2011

  30. [38]

    Deep lstm networks for online chinese handwriting recognition,

    L. Sun, T. Su, C. Liu, and R. Wang, “Deep lstm networks for online chinese handwriting recognition,” in 2016 15th international conference on frontiers in handwriting recognition (icfhr) . IEEE, 2016, pp. 271– 276

Pith tools

Reviewed August 8, 2026 · model on record in the stance chip above.