Pith. sign in

REVIEW 4 major objections 5 minor 1 cited by

A Signer-Invariant Conformer and Multi-Scale Fusion Transformer for Continuous Sign Language Recognition

T0 review · 4 major / 5 minor · reviewed 2026-08-05 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read Two dedicated pose-based networks set new records on the Isharah-1000 sign language benchmark: a conformer hits 13.07% WER on new signers, a multi-scale fusion transformer hits 47.78% on unseen sentences.

desk verdict Solid applied CSLR paper with code and new benchmark numbers, but the SOTA claim is shaky because all comparisons are cross-protocol; the architectures' own contribution isn't isolated. read the letter →

arxiv 2508.09372 v1 pith:S7KIYSC4 submitted 2025-08-12 cs.CV cs.AIcs.IRcs.LG

Md Rezwanul Haque , Md. Milon Islam , S M Taslim Uddin Raju , Fakhri Karray This is my paper
classification cs.CVcs.AIcs.IRcs.LG
keywords continuoussignlanguagerecognitionpose-basedconformertransformersignerindependenceunseensentencegeneralizationworderrorrateIsharah-1000
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

The paper argues that continuous sign language recognition improves when the problem is split into two failure modes and each gets its own pose-based network. For recognizing signs produced by signers never seen in training, it builds a Signer-Invariant Conformer that blends convolution with multi-head self-attention; for recognizing sentence structures never seen in training, it builds a Multi-Scale Fusion Transformer with a dual-path temporal encoder that fuses full-resolution and downsampled motion features. Both models run on 86 2D skeletal keypoints per frame and are trained end-to-end with connectionist temporal classification, so no frame-level gloss alignments are required. On Isharah-1000, the conformer reports a 13.07% word error rate on the signer-independent test set versus 26.6% for the previous best published result, and the transformer reports 47.78% on the unseen-sentence test set versus 48.0% previously. The paper's hypothesis is that task-specific architectural designs, not larger or more generic models, are what move the needle in continuous sign language recognition.

What carries the argument

The two load-bearing mechanisms are the conformer block and the dual-path temporal encoder. The conformer block interleaves multi-head self-attention with a convolution module (depthwise 1D convolution plus gated linear unit) so one stack sees both global context and local gesture shape; this is what the paper credits for signer-agnostic features. The dual-path encoder runs the same input through a full-resolution convolutional path and a max-pooled half-resolution path, concatenates them, and feeds the fused multi-scale sequence to a transformer; this multi-scale view is what the paper credits for handling previously unseen grammatical compositions.

What would settle it

Take the released code and corrupt the hand and finger keypoints (a subset of the 86) in a random 20% of test frames—drop them or replace them with linear-interpolation noise—then measure WER on the signer-independent split. If the 13.07% result degrades sharply, the signer-invariance claim depends on near-perfect upstream keypoints; if it barely moves, the conformer has learned genuine signer-agnostic invariance. A complementary check is to run the same model on keypoints produced by a different pose estimator and compare WER.

Watch

Extended reading notes

Core claim

Both models consume a per-frame vector of 86 2D keypoints covering pose, hands, and face, normalized to the signer's torso and trained end-to-end with connectionist temporal classification, so no frame-level gloss alignment is needed. The signer-independent network passes these keypoints through a convolutional temporal encoder, adds sinusoidal positional encodings, and stacks conformer blocks whose self-attention captures global co-articulation while depthwise convolutions with gated linear units capture local gestural detail. The unseen-sentence network instead first applies a joint attention layer whose queries and keys come from the raw input while the value comes from a later context-aw

Load-bearing premise

The whole pipeline assumes that the 86 keypoints extracted from video are accurate and carry enough information about hand shape, face, and body motion to recognize continuous signing; the models never see the raw RGB frames, so any keypoint error from occlusion, blur, or poor lighting flows straight into the gloss predictions.

Editorial extensions

If this is right

  • If the SI results hold, pose-only keypoints plus a conformer encoder are enough to cut signer-independent sign recognition errors by roughly half compared with the best result previously reported on Isharah-1000.
  • On the US task, the gain over the previous published best is small (0.22 absolute points), so the paper's own results imply that understanding novel sentence structure remains far harder than signer generalization.
  • A simple CNN-BiLSTM baseline outperformed all published methods on the SI test set, indicating that the convolutional feature front-end, not the attention stack alone, carries much of the recognition load.
  • CTC training on keypoint sequences works without gloss-to-frame alignment, keeping the pipeline simpler than methods that require temporal segmentation.
  • Mamba-Sign, the strongest baseline on the US dev set, suggests state-space sequence models deserve attention for the linguistic-generalization half of CSLR.

Reading between the lines

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

  • The paper does not test cross-dataset transfer; a natural next step would be to fine-tune on Isharah and evaluate on a different sign-language corpus to see whether the signer-agnostic representation is language-specific or general.
  • Because RGB frames are never seen, the models' ceiling is keypoint fidelity; the authors do not explore noisy-keypoint augmentation (e.g., dropping or jittering hand joints during training), which could either harden the models against occlusion or reveal the remaining headroom.
  • The near-flat US improvement suggests the remaining errors may be linguistic rather than visual; integrating an explicit gloss-sequence language model or shifting to sign-language translation would test that split.
  • A unified multi-task model with a shared pose encoder and two heads is tempting, but the paper's evidence implies the two objectives may not share features cleanly, since the same baselines that excel on SI fail on US and vice versa.
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 / 5 minor

Summary. The paper addresses continuous sign language recognition (CSLR) on the Isharah-1000 dataset using pose-based skeleton keypoints. Two task-specific architectures are proposed: a Signer-Invariant Conformer for the signer-independent (SI) challenge and a Multi-Scale Fusion Transformer with a joint-attention mechanism and a dual-path temporal encoder for the unseen-sentences (US) challenge. Both are trained with CTC loss. The authors report a test WER of 13.07% on the SI task and 47.78% on the US task, claiming new state-of-the-art results relative to published numbers from the Isharah benchmark. They also implement several baselines and report placements in the SignEval 2025 challenge.

Significance. If the reported results are reproducible and properly controlled, the work would provide evidence that pose-based, task-specific transformer variants can substantially improve CSLR on a challenging real-world dataset. The paper includes a code link and evaluates multiple baselines, which is useful for the community. The proposed architectures are reasonable and the SI result is striking. However, the central 'new SOTA' claim is currently supported only by an uncontrolled comparison against published numbers from a different pipeline, and the lack of ablations and error bars leaves the source of the improvement unclear.

major comments (4)
  1. [§4.5.1, Tables 1–2] The 'new SOTA' claim rests on an uncontrolled cross-protocol comparison. The published baselines are quoted from [7], while every model in this paper consumes 86 keypoints from an unspecified pose estimator (§3.1). The authors' own CNN-BiLSTM baseline (22.62% test WER) already beats the published SOTA (26.6%), so the pose front-end may explain a large part of the improvement. Without re-running prior methods under the same keypoint pipeline, the reported margins cannot be attributed to the proposed architectures.
  2. [§3.2–§3.3, §4.5] No ablation isolates the proposed components. The paper claims that the conformer, joint attention, and dual-path encoder are responsible for the gains, but §4 reports only full-model results. There is no variant without joint attention, without the auxiliary path, or with a single-scale encoder. Consequently, the empirical support for the architectural novelty is missing.
  3. [Table 2, §4.2] No uncertainty quantification is provided. The US result is 47.78% vs. 48.0% for SMKD; without multiple seeds, error bars, or significance tests, this margin is within plausible run-to-run variance. The SI margin is larger, but seed variance should still be reported. Training epochs are given as a range (100–675), which is too coarse for reproduction.
  4. [Eq. (3), §3.3.1] The joint-attention mechanism is not fully specified. The text defines H_att as 'the high-level contextual output from the attention weights block', but no equation or block description explains how H_att is obtained or whether it depends on A_joint. As written, the formula and Figure 2 leave the computation ambiguous, creating a reproducibility gap for a claimed contribution.
minor comments (5)
  1. [Abstract] The abstract says the SI result is 'a reduction of 13.53% from the state-of-the-art'. This is an absolute percentage-point reduction (26.6 – 13.07); the relative reduction is about 50.9%. Please state this precisely.
  2. [§3.1] The pose estimator is never named or versioned. The paper should identify the keypoint extraction framework and, ideally, report its keypoint accuracy on this dataset, since the entire pipeline depends on it.
  3. [§1, Introduction] There is an incomplete sentence: 'large-scale, richly annotated CSLR data [2].' The sentence appears to be missing a verb or continuation.
  4. [Figure 2] The diagram labels around the joint-attention block are unclear (e.g., 'Attention Weights Value (V)'). The flow of Q, K, V should be redrawn or clarified to match Eq. (3).
  5. [Table 2] Several baselines show development WER but no test WER; the reason for missing test values should be stated. Also, the test column for the proposed Multi-Scale Fusion Transformer should be bolded consistently with the table caption.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the paper's contributions are empirical and the architectures are defined as feedforward compositions without self-referential equations.

full rationale

The paper is an empirical systems paper. The proposed models are defined as explicit feedforward compositions: raw keypoints X' are projected, processed by temporal encoders, conformer/transformer blocks, and a linear classifier, trained with CTC loss. The only potentially suspicious equation, Eq. (3), Ajoint = Attention(X'W_Q, X'W_K, H_att W_V), uses H_att as the output of a separate 'attention weights block' (Figure 2) rather than the network's final output; the paper does not define H_att in terms of Ajoint, so no self-referential cycle is established. Reported results are test-set WER after standard training and validation, not fitted constants recycled as predictions. Prior benchmark numbers are imported from [7] rather than re-run on the same pose-keypoint protocol, which is a legitimate comparability and attribution concern—especially since the authors' own CNN-BiLSTM baseline already surpasses the published SOTA—but this is an evaluation-validity limitation, not circularity. The authors explicitly acknowledge the dependence on the upstream keypoint extractor in Section 5. No load-bearing self-citation, imported uniqueness theorem, or renaming of known results is present. The derivation chain, to the extent that one exists, is self-contained; the SOTA claim depends on external benchmark comparison but does not reduce to the paper's own inputs.

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

The central claim rests on a handful of domain assumptions about the pose representation, the benchmark splits, and the comparability of previously published results. No new physical or conceptual entities are introduced. The only fitted values are network weights trained by gradient descent, plus the hyperparameters listed above, but many architectural dimensions are omitted from the paper.

free parameters (5)
  • learning_rate = 1e-4
    Reported in Section 4.2; chosen by hand, central to convergence of both models.
  • batch_size = 16
    Reported in Section 4.2; affects optimization dynamics.
  • epochs = 100-675
    Section 4.2 reports a range; exact stopping criterion for each model is not specified, implying early selection on dev WER.
  • dropout = 0.1 (conformer), 0.3 (classifier head)
    Mentioned in Figure captions and Section 3.3.4; chosen by hand.
  • model_dimensions
    Number of conformer/transformer blocks, hidden sizes, attention heads, and kernel sizes are not reported, yet they determine the reported WER values.
assumptions (4)
  • domain assumption CTC loss is a valid weak supervision for CSLR without frame-level gloss alignment.
    Invoked in Section 3.2.3 and 3.3.4; the models rely on CTC to learn alignments.
  • domain assumption The 86 keypoints from the pose estimator provide sufficient information for sign recognition.
    Stated in Section 3.1; the models never see raw RGB frames, so this assumption is load-bearing.
  • domain assumption The Isharah-1000 official splits correctly operationalize signer-independence and unseen-sentence generalization.
    Section 4.1 describes the splits but does not justify their construct validity beyond the original dataset paper.
  • domain assumption Prior published results are comparable to this paper's results despite different preprocessing and training details.
    Tables 1 and 2 quote prior results from [7] without re-running under the same pose pipeline; this assumes comparability.

how reviews work

0 comments
Cite this review

Pith. "Pith review of A Signer-Invariant Conformer and Multi-Scale Fusion Transformer for Continuous Sign Language Recognition." pith.science (2026). https://pith.science/paper/S7KIYSC4

@misc{pith2026250809372,
  author       = {Pith},
  title        = {Pith review of: A Signer-Invariant Conformer and Multi-Scale Fusion Transformer for Continuous Sign Language Recognition},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/S7KIYSC4}},
  note         = {Machine review of arXiv:2508.09372}
}
read the original abstract

Continuous Sign Language Recognition (CSLR) faces multiple challenges, including significant inter-signer variability and poor generalization to novel sentence structures. Traditional solutions frequently fail to handle these issues efficiently. For overcoming these constraints, we propose a dual-architecture framework. For the Signer-Independent (SI) challenge, we propose a Signer-Invariant Conformer that combines convolutions with multi-head self-attention to learn robust, signer-agnostic representations from pose-based skeletal keypoints. For the Unseen-Sentences (US) task, we designed a Multi-Scale Fusion Transformer with a novel dual-path temporal encoder that captures both fine-grained posture dynamics, enabling the model's ability to comprehend novel grammatical compositions. Experiments on the challenging Isharah-1000 dataset establish a new standard for both CSLR benchmarks. The proposed conformer architecture achieves a Word Error Rate (WER) of 13.07% on the SI challenge, a reduction of 13.53% from the state-of-the-art. On the US task, the transformer model scores a WER of 47.78%, surpassing previous work. In the SignEval 2025 CSLR challenge, our team placed 2nd in the US task and 4th in the SI task, demonstrating the performance of these models. The findings validate our key hypothesis: that developing task-specific networks designed for the particular challenges of CSLR leads to considerable performance improvements and establishes a new baseline for further research. The source code is available at: https://github.com/rezwanh001/MSLR-Pose86K-CSLR-Isharah.

Figures

Figures reproduced from arXiv: 2508.09372 by the authors.

Figure 1
Figure 1. Signer-Invariant Conformer: our proposed architecture for signer-independent CSLR begins by extracting pose keypoints from video frames. An initial temporal encoder, composed of convolutional layers, learns local features from this pose sequence. The core of the model consists of conformer blocks that capture global context with multi-head self-attention and extract local patterns using convolution. Positional encod… view at source ↗
Figure 2
Figure 2. Multi-Scale Fusion Transformer: an overview of the proposed architecture for the unseen sentences CSLR task. The network first uses a pose estimator to retrieve keypoint data. A joint attention module is then immediately applied to weigh the importance of features at the frame level. The generated features are then processed by a temporal encoder with a dual-path design: a main block records fine-grained temporal dy… view at source ↗

Discussion (0). Sign in to comment.

Forward citations

Cited by 1 Pith paper

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

  1. Pseudo Empirical Likelihood Inference for Non-Probability Survey Samples

    stat.ME 2025-08 unverdicted novelty 5.0 of 10

    Pseudo empirical likelihood confidence intervals for non-probability survey samples are made range-respecting and data-driven, with simulation support for binary responses.

Reference graph

Works this paper leans on

39 extracted references · 33 canonical work pages · cited by 1 Pith paper

  1. [7]

    Isharah: A Large-Scale Multi-Scene Dataset for Continuous Sign Language Recognition

    Sarah Alyami, Hamzah Luqman, Sadam Al-Azani, Maad Alowaifeer, Yazeed Alharbi, and Yaser Alonaizan. Isharah: A large-scale multi-scene dataset for continuous sign lan- guage recognition. arXiv:2506.03615, 2025. 2, 6, 7, 8

    work page Pith review arXiv 2025

  2. [1]

    https : / / www

    Hearing loss statistics. https : / / www . who . int / newsroom / fact - sheets / detail / deafness - and-hearing-loss. Accessed: 2025-06-30. 1

    work page 2025

  3. [2]

    A comprehen- sive study on deep learning-based methods for sign language recognition

    Nikolas Adaloglou, Theocharis Chatzis, Ilias Papastratis, Andreas Stergioulas, Georgios Th Papadopoulos, Vassia Zacharopoulou, George J Xydopoulos, Klimnis Atzakas, Dimitris Papazachariou, and Petros Daras. A comprehen- sive study on deep learning-based methods for sign language recognition. IEEE Transactions on Multimedia , 24:1750– 1762, 2021. 1

    work page 2021

  4. [3]

    Slow- fast network for continuous sign language recognition

    Junseok Ahn, Youngjoon Jang, and Joon Son Chung. Slow- fast network for continuous sign language recognition. In ICASSP 2024-2024 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP) , pages 3920–3924. IEEE, 2024. 7, 8

    work page 2024

  5. [4]

    Understanding vision-based continuous sign language recognition

    Neena Aloysius and M Geetha. Understanding vision-based continuous sign language recognition. Multimedia Tools and Applications, 79(31):22177–22209, 2020. 1, 2 8

    work page 2020

  6. [5]

    Swin-mstp: Swin trans- former with multi-scale temporal perception for continuous sign language recognition

    Sarah Alyami and Hamzah Luqman. Swin-mstp: Swin trans- former with multi-scale temporal perception for continuous sign language recognition. Neurocomputing, 617:129015,

    work page

  7. [6]

    Reviewing 25 years of continuous sign lan- guage recognition research: Advances, challenges, and prospects

    Sarah Alyami, Hamzah Luqman, and Mohammad Ham- moudeh. Reviewing 25 years of continuous sign lan- guage recognition research: Advances, challenges, and prospects. Information Processing & Management , 61(5): 103774, 2024. 1, 6

    work page 2024

  8. [8]

    Arabert: Transformer-based model for arabic language understand- ing

    Wissam Antoun, Fady Baly, and Hazem Hajj. Arabert: Transformer-based model for arabic language understand- ing. arXiv:2003.00104, 2020. 7

    arXiv 2003

Show all 39 references
  1. [9]

    A signer independent sign language recognition with co-articulation elimination from live videos: an indian scenario

    PK Athira, CJ Sruthi, and A Lijiya. A signer independent sign language recognition with co-articulation elimination from live videos: an indian scenario. Journal of King Saud University-Computer and Information Sciences , 34(3):771– 781, 2022. 1, 4

    2022

  2. [10]

    Subunets: End-to-end hand shape and continuous sign language recognition

    Necati Cihan Camgoz, Simon Hadfield, Oscar Koller, and Richard Bowden. Subunets: End-to-end hand shape and continuous sign language recognition. In Proceedings of the IEEE International Conference on Computer Vision , pages 3056–3065, 2017. 3

    2017

  3. [11]

    Unsupervised cross-lingual representation learn- ing at scale

    Alexis Conneau, Kartikay Khandelwal, Naman Goyal, Vishrav Chaudhary, Guillaume Wenzek, Francisco Guzm´an, Edouard Grave, Myle Ott, Luke Zettlemoyer, and Veselin Stoyanov. Unsupervised cross-lingual representation learn- ing at scale. arXiv:1911.02116, 2019. 7

    1911 arXiv

  4. [12]

    A deep neu- ral framework for continuous sign language recognition by iterative training

    Runpeng Cui, Hu Liu, and Changshui Zhang. A deep neu- ral framework for continuous sign language recognition by iterative training. IEEE Transactions on Multimedia, 21(7): 1880–1891, 2019. 3

    2019

  5. [13]

    Spatial–temporal transformer for end-to-end sign language recognition

    Zhenchao Cui, Wenbo Zhang, Zhaoxin Li, and Zhaoqi Wang. Spatial–temporal transformer for end-to-end sign language recognition. Complex & Intelligent Systems , 9(4):4645– 4656, 2023. 2

    2023

  6. [14]

    A survey on continuous sign language recognition

    Maria De Sisto, Gaetano Sequino, and Angelo Ciaramella. A survey on continuous sign language recognition. ACM Computing Surveys, 55(10):1–38, 2023. 1

    2023

  7. [15]

    A comprehensive survey and taxonomy of sign language research.Engineering Applications of Artificial Intelligence, 114:105198, 2022

    El-Sayed M El-Alfy and Hamzah Luqman. A comprehensive survey and taxonomy of sign language research.Engineering Applications of Artificial Intelligence, 114:105198, 2022. 1

    2022

  8. [16]

    Extensions of the sign language recognition and translation corpus rwth-phoenix-weather

    Jens Forster, Christoph Schmidt, Oscar Koller, Martin Bell- gardt, and Hermann Ney. Extensions of the sign language recognition and translation corpus rwth-phoenix-weather. In LREC, pages 1911–1916, 2014. 2

    1911

  9. [17]

    Mamba: Linear-time sequence mod- eling with selective state spaces

    Albert Gu and Tri Dao. Mamba: Linear-time sequence mod- eling with selective state spaces. arXiv:2312.00752, 2023. 7

    2023 arXiv

  10. [18]

    Conformer: Convolution-augmented transformer for speech recognition

    Anmol Gulati, James Qin, Chung-Cheng Chiu, Niki Par- mar, Yu Zhang, Jiahui Yu, Wei Han, Shibo Wang, Zhengdong Zhang, Yonghui Wu, et al. Conformer: Convolution-augmented transformer for speech recognition. arXiv:2005.08100, 2020. 4, 7

    2005 arXiv

  11. [19]

    Self-mutual dis- tillation learning for continuous sign language recognition

    Aiming Hao, Yuecong Min, and Xilin Chen. Self-mutual dis- tillation learning for continuous sign language recognition. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11303–11312, 2021. 7, 8

    2021

  12. [20]

    Long short-term memory

    Sepp Hochreiter and J ¨urgen Schmidhuber. Long short-term memory. Neural Computation, 9(8):1735–1780, 1997. 7

    1997

  13. [21]

    Tempo- ral lift pooling for continuous sign language recognition

    Lianyu Hu, Liqing Gao, Zekang Liu, and Wei Feng. Tempo- ral lift pooling for continuous sign language recognition. In European Conference on Computer Vision, pages 511–527. Springer, 2022. 7, 8

    2022

  14. [22]

    Con- tinuous sign language recognition with correlation network

    Lianyu Hu, Liqing Gao, Zekang Liu, and Wei Feng. Con- tinuous sign language recognition with correlation network. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2529–2539, 2023. 7, 8

    2023

  15. [23]

    Self- emphasizing network for continuous sign language recogni- tion

    Lianyu Hu, Liqing Gao, Zekang Liu, and Wei Feng. Self- emphasizing network for continuous sign language recogni- tion. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 854–862, 2023. 7, 8

    2023

  16. [24]

    Video-based sign language recognition without temporal segmentation

    Jie Huang, Wengang Zhou, Qilin Zhang, Houqiang Li, and Weiping Li. Video-based sign language recognition without temporal segmentation. In Proceedings of the AAAI Confer- ence on Artificial Intelligence, 2018. 2, 3

    2018

  17. [25]

    Theruslan: Database of rus- sian sign language

    Ildar Kagirov, Denis Ivanko, Dmitry Ryumin, Alexander Axyonov, and Alexey Karpov. Theruslan: Database of rus- sian sign language. In Proceedings of the Twelfth Language Resources and Evaluation Conference , pages 6079–6085,

  18. [26]

    Temporal convolutional networks: A unified approach to ac- tion segmentation

    Colin Lea, Rene Vidal, Austin Reiter, and Gregory D Hager. Temporal convolutional networks: A unified approach to ac- tion segmentation. In Computer vision–ECCV 2016 work- shops: Amsterdam, the Netherlands, October 8-10 and 15- 16, 2016, proceedings, part III 14 , pages 47–54....

    2016

  19. [27]

    Word-level deep sign language recognition from video: A new large-scale dataset and methods comparison

    Dongxu Li, Cristian Rodriguez, Xin Yu, and Hongdong Li. Word-level deep sign language recognition from video: A new large-scale dataset and methods comparison. In Pro- ceedings of the IEEE/CVF Winter Conference on Applica- tions of Computer Vision, pages 1459–1469, 2020. 1

    2020

  20. [28]

    Multi-view spatial-temporal network for continuous sign language recognition

    Ronghui Li and Lu Meng. Multi-view spatial-temporal network for continuous sign language recognition. arXiv:2204.08747, 2022. 2, 3

    2022 arXiv

  21. [29]

    Automatic transla- tion of arabic text-to-arabic sign language

    Hamzah Luqman and Sabri A Mahmoud. Automatic transla- tion of arabic text-to-arabic sign language. Universal Access in the Information Society, 18(4):939–951, 2019. 1

    2019

  22. [30]

    The signeval 2025 challenge at the iccv mul- timodal sign language recognition workshop: Results and discussion

    Hamzah Luqman, Raffaele Mineo, Murtadha Aljubran, Ahmed Abul Hasanaath, Amelia Sorrenti, Sarah Alyami, Sadam Al-Azani, Maad Alowaifeer, JiHwan Moon, Vaclav Javorek, Tomas Zelezny, Marek Hruz, Gaia Caligiore, Sil- vio Giancola, Senya Polikovsky, Motaz Alfarraj, Sabina Fontana, ...

    2025

  23. [31]

    Visual alignment constraint for continuous sign language recognition

    Yuecong Min, Aiming Hao, Xiujuan Chai, and Xilin Chen. Visual alignment constraint for continuous sign language recognition. In proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11542–11551, 2021. 7, 8

    2021

  24. [32]

    Fluentsigners-50: A signer independent benchmark dataset for sign language processing

    Medet Mukushev, Aidyn Ubingazhibov, Aigerim Kydyr- bekova, Alfarabi Imashev, Vadim Kimmelman, and Anara Sandygulova. Fluentsigners-50: A signer independent benchmark dataset for sign language processing. PLOS One, 17(9):e0273649, 2022. 2

    2022

  25. [33]

    Distilbert, a distilled version of bert: smaller, faster, cheaper and lighter

    Victor Sanh, Lysandre Debut, Julien Chaumond, and Thomas Wolf. Distilbert, a distilled version of bert: smaller, faster, cheaper and lighter. arXiv:1910.01108, 2019. 7

    1910 arXiv

  26. [34]

    Karsl: Arabic sign language database

    Ala Addin I Sidig, Hamzah Luqman, Sabri Mahmoud, and Mohamed Mohandes. Karsl: Arabic sign language database. ACM Transactions on Asian and Low-Resource Language Information Processing (TALLIP), 20(1):1–19, 2021. 1

    2021

  27. [35]

    Llama 2: Open foundation and fine-tuned chat models.arXiv preprint arXiv:2307.09288, 2023

    Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, et al. Llama 2: Open foundation and fine-tuned chat models.arXiv preprint arXiv:2307.09288, 2023. 7

    2023 arXiv

  28. [36]

    Signum database: Video corpus for signer-independent continuous sign language recognition

    Ulrich von Agris and Karl-Friedrich Kraiss. Signum database: Video corpus for signer-independent continuous sign language recognition. In 4th Workshop on the Repre- sentation and Processing of Sign Languages: Corpora and Sign Language Technologies, pages 243–246, 2010. 2

    2010

  29. [37]

    Spatial tempo- ral graph convolutional networks for skeleton-based action recognition

    Sijie Yan, Yuanjun Xiong, and Dahua Lin. Spatial tempo- ral graph convolutional networks for skeleton-based action recognition. In Proceedings of the AAAI Conference on Ar- tificial Intelligence, 2018. 3, 7

    2018

  30. [38]

    Wisign: Ubiquitous american sign language recognition using com- mercial wi-fi devices

    Lei Zhang, Yixiang Zhang, and Xiaolong Zheng. Wisign: Ubiquitous american sign language recognition using com- mercial wi-fi devices. ACM Transactions on Intelligent Sys- tems and Technology, 11(3), 2020. 3

    2020

  31. [39]

    Spatial-temporal multi-cue network for continuous sign lan- guage recognition

    Hao Zhou, Wengang Zhou, Yun Zhou, and Houqiang Li. Spatial-temporal multi-cue network for continuous sign lan- guage recognition. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 13009–13016, 2020. 3 10

    2020

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.