Pith. sign in

REVIEW 4 major objections 5 minor 88 references

Using Sign Language Production as Data Augmentation to enhance Sign Language Translation

T0 review · 4 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read Pre-training a pose-based sign language translator on stitched synthetic skeleton sequences, then fine-tuning on real video, lifts BLEU-1 on PHOENIX14T from 32.41 to 38.61 (a 19.1% relative gain).

desk verdict The skeleton-stitching pre-training gain is consistent across many configurations, but no error bars and test-set hyperparameter selection make the headline 19% unreliable. read the letter →

arxiv 2506.09643 v1 pith:MGB75SDZ submitted 2025-06-11 cs.CL cs.CV

Harry Walsh , Maksym Ivashechkin , Richard Bowden This is my paper
classification cs.CLcs.CV
keywords SignLanguageTranslationProductionDataAugmentationStitchingSkeletonPoseGaussianSplattingGenerativeAdversarialNetworkPHOENIX14T
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

Sign language translation models are held back by tiny datasets: PHOENIX14T has only about 9,000 sentences. This paper argues that three Sign Language Production techniques can manufacture extra training data for them. The central result is that stitching isolated signs into synthetic skeleton sequences and pre-training a pose-based translator on them, before fine-tuning on real data, raises BLEU-1 from 32.41 to 37.86, and to 38.61 when the synthetic sequences are speeded up by 1.5. The paper also reports that rendering synthetic signers with SignSplat improves video-based translation BLEU-4 scores, while SignGAN's noisier outputs help only some metrics. If these gains hold, synthetic SLP data becomes a practical remedy for low-resource sign languages, delivering more than 15% improvement on every metric tested.

What carries the argument

The load-bearing object is the sign-stitching pipeline: a dictionary of gloss-to-joint-angle sequences, duration prediction, linear-interpolated sign transitions, and a Butterworth low-pass filter that shapes the synthetic motion to resemble real signing. It produces the pre-training corpus and is the mechanism that lets the model see lexical variants, speed variations, and order permutations before touching real data.

What would settle it

Remove the Butterworth low-pass filter, or replace the interpolated transitions with plain concatenation, and re-run the pre-training; if BLEU-1 on the PHOENIX14T test set stops improving over the 32.41 baseline or drops, the motion-filtering step is what carries the transfer and the general claim that stitched sequences suffice would be falsified.

Watch

Extended reading notes

Core claim

The paper claims that synthetic skeleton sequences generated by sign stitching are a viable pre-training corpus for pose-to-text sign language translation. The authors construct a dictionary of 7,206 isolated DGS signs, resample each to a target duration, concatenate them with linear-interpolated transitions, and low-pass filter the whole sequence to mimic natural motion. Pre-training the Sign Language Transformer on this stitched data and then fine-tuning on the real PHOENIX14T data outperforms the same model trained only on real data across all BLEU and ROUGE metrics, with BLEU-1 rising from 32.41 to 38.61 at a 1.5 duration scale. They additionally show that photo-realistic appearance rendering, SignSplat in particular, can supplement video-based translation, improving BLEU-4 on test and dev sets, while SignGAN's artefacts limit its contribution.

Load-bearing premise

The whole approach rests on the assumption that skeleton sequences stitched from isolated signs and low-pass filtered are close enough in motion statistics to real continuous signing that a model pre-trained on them transfers to real data, and the paper evaluates this only on one dataset and one architecture.

Editorial extensions

If this is right

  • Every BLEU and ROUGE metric on PHOENIX14T improves by at least about 15% when the pose-based translator is pre-trained on stitched skeletons instead of trained from scratch.
  • Varying the synthetic sequence duration by a factor of 1.5 yields the strongest test-set result (BLEU-1 of 38.61), suggesting speed variation acts as a regularizer for the pose-to-text model.
  • Permuting up to 3 signs in a sequence slightly improves higher-n-gram BLEU scores on the dev set, while larger permutations hurt, so grammatical order carries information the model relies on.
  • Photo-realistic appearance augmentation from SignSplat improves video-to-text BLEU-4 scores, while SignGAN's data helps only a subset of metrics because of rendering artefacts.

Reading between the lines

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

  • Beyond the paper, the stitched-skeleton recipe is in principle architecture-agnostic: any pose-to-text model consuming the same keypoint representation might show a similar gain, but since only one architecture is tested, transfer across encoders remains an open question.
  • Because the stitched data is built from a dictionary of isolated signs, the approach could extend to any sign language that has both a sign dictionary and a text-aligned corpus; gains may scale with how well the dictionary covers the target dataset's vocabulary.
  • The 1.5 duration-scale result hints that making synthetic data harder (faster) helps; an untested corollary is that a curriculum over duration scales, from slow to fast, could push gains further.
  • Combining novel skeleton motion with the SignSplat appearance model to render multi-view synthetic video would test whether the two augmentation axes are additive; the paper names this direction but does not run it.
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. This paper proposes using three sign language production (SLP) methods as data augmentation for sign language translation (SLT): sign stitching to synthesize continuous skeleton sequences, and two photorealistic generators, SignGAN and SignSplat, to vary signer appearance. The main experiment pre-trains a pose-to-text transformer (Sign Language Transformers) on stitched synthetic skeletons and fine-tunes on PHOENIX14T, reporting BLEU-1 gains from 32.41 to 37.86, and to 38.61 when the sequence duration is scaled by 1.5. Video-to-text experiments with GF-SLT show mixed but modest improvements from appearance augmentation, with SignSplat-augmented training giving the best BLEU-4. The paper concludes that synthetic skeleton pre-training is beneficial and that visual appearance augmentation is promising but more limited.

Significance. If the reported improvements are robust, this is a useful contribution to low-resource sign language translation: it is, to the authors' knowledge, the first proposal to use generated synthetic skeleton sequences for SLT pre-training, and it systematically compares multiple augmentation strategies on a standard benchmark. The authors also provide a useful negative result: direct joint training on real and stitched data is not beneficial, while pre-training followed by fine-tuning is. The main weakness is that the headline claim rests on single-run point estimates with hyperparameters selected from the test set, so the magnitude and even the existence of the effect are not yet established beyond run-to-run variation.

major comments (4)
  1. [Section 4.2.1 (Table 1)] The headline result is supported only by single unseeded runs for each configuration, and the best configuration (duration scale 1.5) is selected from the test set. No variance, confidence intervals, or significance tests are reported anywhere in Section 4. Since the PHOENIX14T test set is small and transformer training on such data can exhibit nontrivial run-to-run variance, the 16.8% and 19.1% relative BLEU-1 gains are not yet established as reproducible. Please report multiple seeds (at least 3-5) for the baseline, the best configuration, and the main comparison rows, or use the dev set for hyperparameter selection and restrict the test set to the final chosen model.
  2. [Section 4.2.1 (text after Table 1.a)] The text claims that pre-training "provides at least 15% improvement for all metrics," but Table 1.a shows ROUGE on the test set improving from 32.96 to 37.61, which is a 14.1% relative gain, not at least 15%. Please correct the claim or the reported numbers.
  3. [Section 4.2.1 (Table 1.c and surrounding text)] The text states that the best performance comes from "increasing the speed by 1.5 times" and "reducing the total number of frames," while the table labels the configuration as "Duration scale: 1.5." If the scale is applied to per-gloss durations, a scale of 1.5 produces longer sequences with more frames, i.e., slower signing. This internal inconsistency affects the interpretation of the main augmentation mechanism and should be clarified.
  4. [Section 4.2.2 (Table 2)] The statement that SignSplat augmentation increases BLEU-4 by 0.49 on the test set and 0.68 on the dev set does not match Table 2. Comparing "PHIX + Splat" with the VLP-pretrained baseline "PHIX (VLP PreTrain)" gives +0.44 test and +0.35 dev, while comparing with the un-pretrained "PHIX" row gives +2.26 and +1.85. Please state the comparison baseline explicitly and correct the numbers.
minor comments (5)
  1. [Section 3.1.1] The word-embedding substitution for missing dictionary signs is not specified: which embedding model is used, what similarity threshold is applied, and whether substituted signs were validated. Please provide these details.
  2. [Section 3.2.1 / Section 4.1] The pre-training protocol for the P2T model is under-specified: the number of synthetic sequences used, the pre-training epoch count, the learning rate schedule, and the fine-tuning procedure should be stated to allow reproduction.
  3. [Section 4.2.1 (Table 1.a)] The sentence describing the stitched-only row as "impressive" is subjective; consider instead quantifying the result (e.g., BLEU-1 is about half the baseline) or removing the evaluative language.
  4. [Section 4.2.1 (last paragraph)] The statement "Given only a 70% sign overlap in the dev and test split" is unclear; please define whether the overlap is at the gloss-vocabulary level, sentence level, or sign-occurrence level, and explain how it motivates the speed-bias interpretation.
  5. [Throughout] There are minor typos and notation inconsistencies, including "an BLEU-1" in Section 4.2.1, mixed "signGAN"/"SignGAN" spellings, and inconsistent capitalization of "Sign Language Transformer." Please proofread.

Circularity Check

0 steps flagged · score 0.0 of 10

No derivation-level circularity: the central claim is an empirical benchmark comparison, and self-citations are used only as tooling and baselines, not as load-bearing evidence.

full rationale

The paper's claimed contribution is empirical: it uses SLP-generated skeletons and videos as augmentation and measures BLEU/ROUGE on the held-out PHOENIX14T test split. Tracing the chain from Section 3.1.1 to Section 4.2.1, the synthetic sequences are constructed from a dictionary of isolated signs, duration estimates, linear-interpolated transitions, and Butterworth filtering. None of these steps is defined in terms of the reported BLEU score, and no model parameter is fitted to the test set. The baseline and augmented runs share the same architecture and training protocol, so the comparison is self-contained against an external benchmark. Citations to prior work by the same authors ([35], [36], [64], [76]) supply generators and the pose pipeline, but these are used as off-the-shelf tools rather than as evidence for the headline improvement; the improvement itself is a direct measurement on an external test set. The selection of the duration-scale 1.5 configuration based on test-set performance and the absence of seed variance are legitimate statistical-validity concerns, but they are not circularity: the central qualitative result (stitched pre-training beats GT-only for every reported BLEU-1 configuration) does not reduce, by construction or by self-citation, to its own input. No equation or fitted parameter can be exhibited that makes the reported prediction equivalent to the inputs by definition, so the appropriate finding is no significant circularity.

Assumptions & free parameters 6 free parameters · 5 assumptions · 0 invented entities

The paper rests on several domain assumptions about the fidelity of synthetic signing and the representativeness of the PHOENIX14T benchmark; no new physical entities are introduced.

free parameters (6)
  • Butterworth low-pass cutoff frequency C = unspecified (per-sequence)
    Used in motion filtering (Sec 3.1.1); no procedure given for setting it; affects naturalness of generated sequences.
  • Duration scale factor = 1.5 (best on test set)
    In Table 1.c, 1.5 gives the highest test BLEU-1; 0.7 is best on dev, so the choice is test-set selected.
  • Number of sign-order permutations N = 0 to 10 tested; 0 or 3 used for best n-grams
    In Table 1.b, N is a hyperparameter; marginal effects.
  • Number of transition frames = determined by boundary distance and velocity bound
    Linear interpolation length is a design choice described qualitatively in Sec 3.1.1.
  • Word embedding model for gloss substitution = unspecified
    Used to pick the nearest dictionary sign; model not named in the paper.
  • SMPL-X optimization hyperparameters (learning rates, iterations) = small learning rate for hands; 20-40 sec per video
    Taken from prior SignSplat paper [36]; not tuned here.
assumptions (5)
  • domain assumption Stitched skeleton sequences are a faithful proxy for real continuous signing motion.
    The whole P2T pre-training relies on this; enters Section 3.1.1 and is not independently verified.
  • domain assumption The dictionary of 7,206 DGS signs covers the PHOENIX14T gloss vocabulary semantics sufficiently for substitution.
    Coverage is claimed, but place names are missing; paraphrases may be semantically off.
  • domain assumption The Sign Language Transformer is a representative P2T architecture for evaluating augmentation benefit.
    Only one P2T and one V2T architecture are tested; generalization to SOTA models is assumed.
  • domain assumption Generated avatar videos preserve the manual features needed for V2T translation.
    SignGAN artifacts (blurring, hand merging) are acknowledged; SignSplat is more robust, but transfer to V2T is still assumed.
  • domain assumption MediaPipe 3D pose and the canonical skeleton normalization do not introduce signer-specific bias that dominates augmentation gains.
    Pose extraction is from [35]; used for both real and generated sequences.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Using Sign Language Production as Data Augmentation to enhance Sign Language Translation." pith.science (2026). https://pith.science/paper/MGB75SDZ

@misc{pith2026250609643,
  author       = {Pith},
  title        = {Pith review of: Using Sign Language Production as Data Augmentation to enhance Sign Language Translation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/MGB75SDZ}},
  note         = {Machine review of arXiv:2506.09643}
}
read the original abstract

Machine learning models fundamentally rely on large quantities of high-quality data. Collecting the necessary data for these models can be challenging due to cost, scarcity, and privacy restrictions. Signed languages are visual languages used by the deaf community and are considered low-resource languages. Sign language datasets are often orders of magnitude smaller than their spoken language counterparts. Sign Language Production is the task of generating sign language videos from spoken language sentences, while Sign Language Translation is the reverse translation task. Here, we propose leveraging recent advancements in Sign Language Production to augment existing sign language datasets and enhance the performance of Sign Language Translation models. For this, we utilize three techniques: a skeleton-based approach to production, sign stitching, and two photo-realistic generative models, SignGAN and SignSplat. We evaluate the effectiveness of these techniques in enhancing the performance of Sign Language Translation models by generating variation in the signer's appearance and the motion of the skeletal data. Our results demonstrate that the proposed methods can effectively augment existing datasets and enhance the performance of Sign Language Translation models by up to 19%, paving the way for more robust and accurate Sign Language Translation systems, even in resource-constrained environments.

Figures

Figures reproduced from arXiv: 2506.09643 by the authors.

Figure 1
Figure 1. A visual overview of the SLT and SLP tasks. The SLT task is the process of translating sign language video into spoken language text. The SLP task is the reverse, generating sign language video from spoken language text. methods that attempt to directly regress a sequence of poses from the spoken language text [63]. Mainly regression to the mean, which causes under-articulated and often incomprehensible signing. For… view at source ↗
Figure 2
Figure 2. An example from the PHOENIX14T dataset, showing left to right: skeleton pose, original video, SignSplat Avatar, and SignGAN Avatar. the augmented data, followed by a second stage of fine-tuning on real data at a lower learning rate. Others propose simultaneously training on both real and augmented data [17]. In this work, we experiment with both of these approaches. 2.2 Sign Language Production SLP is the task of ge… view at source ↗
Figure 3
Figure 3. An example from the PHOENIX14T dataset, showing top to bottom: spoken language, gloss, original video, extracted skeleton, and Stitched sequence. BLEU-2 to 4 scores. Given only a 70% sign overlap in the dev and test split, we suspect speed augmentation might introduce bias for some signs. 4.2.2 Video-to-Text [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: An example from the PHOENIX14T dataset, showing top to bottom: spoken language, gloss of the original video, SignSplat Avatar, and SignGAN Avatar [PITH_FULL_IMAGE:figures/full_fig_p008_4.png]
Figure 5
Figure 5. Figure 5: Examples from the PHOENIX14T dataset, showing top to bottom: the original video, SignSplat Avatar, and Sign￾GAN Avatar. The bottom row shows issues with blurring, hand merging and artefacts in the face. of the GAN architecture and the training data used. As shown in […

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

88 extracted references · 36 canonical work pages

  1. [1]

    Sharif Md Abdullah, Abhijit Paul, Shebuti Rayana, Ahmedul Kabir, and Zarif Masud. 2025. State-of-the-Art Translation of Text-to-Gloss using mBART: A case study of Bangla.arXiv preprint arXiv:2504.02293(2025)

    arXiv 2025

  2. [2]

    Samuel Albanie, Gül Varol, Liliane Momeni, Hannah Bull, Triantafyllos Afouras, Himel Chowdhury, Neil Fox, Bencie Woll, Rob Cooper, Andrew McParland, et al

  3. [3]

    Hasan Algafri, Hamzah Luqman, Sarah Alyami, and Issam Laradji. 2025. SSLR: A Semi-Supervised Learning Method for Isolated Sign Language Recognition. arXiv preprint arXiv:2504.16640(2025)

    work page Pith review arXiv 2025

  4. [4]

    Andrew Bangham, SJ Cox, Ralph Elliott, John RW Glauert, Ian Marshall, Sanja Rankov, and Mark Wells. 2000. Virtual signing: Capture, animation, storage and transmission-an overview of the visicast project. InIEE Seminar on speech and language processing for disabled and elderly people (Ref. No. 2000/025). IET, 6–1

    2000

  5. [5]

    Danielle Bragg, Oscar Koller, Mary Bellard, Larwan Berke, Patrick Boudreault, Annelies Braffort, Naomi Caselli, Matt Huenerfauth, Hernisa Kacorri, Tessa Verhoef, et al. 2019. Sign language recognition, generation, and translation: An interdisciplinary perspective. InProceedings of the 21st international ACM SIGACCESS conference on computers and accessibil...

    2019

  6. [6]

    Stephen Butterworth et al . 1930. On the theory of filter amplifiers.Wireless Engineer7, 6 (1930), 536–541

    1930

  7. [7]

    Necati Cihan Camgoz, Simon Hadfield, Oscar Koller, Hermann Ney, and Richard Bowden. 2018. Neural sign language translation. InProceedings of the IEEE conference on computer vision and pattern recognition. 7784–7793

    2018

  8. [8]

    Necati Cihan Camgoz, Oscar Koller, Simon Hadfield, and Richard Bowden. 2020. Sign language transformers: Joint end-to-end sign language recognition and translation. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition. 10023–10033

    2020

Show all 88 references
  1. [9]

    Yutong Chen, Ronglai Zuo, Fangyun Wei, Yu Wu, Shujie Liu, and Brian Mak. 2022. Two-stream network for sign language recognition and translation.Advances in Neural Information Processing Systems35 (2022), 17043–17056

    2022

  2. [10]

    MMPose Contributors. 2020. OpenMMLab Pose Estimation Toolbox and Bench- mark. https://github.com/open-mmlab/mmpose

    2020

  3. [11]

    Stephen Cox, Michael Lincoln, Judy Tryggvason, Melanie Nakisa, Mark Wells, Marcus Tutt, and Sanja Abbott. 2002. Tessa, a system to aid communication with deaf people. InProceedings of the fifth international ACM conference on Assistive technologies. 205–212

    2002

  4. [12]

    Shengliang Deng, Mi Yan, Songlin Wei, Haixin Ma, Yuxin Yang, Jiayi Chen, Zhiqi Zhang, Taoyu Yang, Xuheng Zhang, Heming Cui, et al . 2025. GraspVLA: a Grasping Foundation Model Pre-trained on Billion-scale Synthetic Action Data. arXiv preprint arXiv:2505.03233(2025)

    2025 arXiv

  5. [13]

    Kenneth W Dunn, Chichen Fu, David Joon Ho, Soonam Lee, Shuo Han, Paul Salama, and Edward J Delp. 2019. DeepSynth: Three-dimensional nuclear seg- mentation of biological images using neural networks trained with synthetic data.Scientific reports9, 1 (2019), 18295

    2019

  6. [14]

    Eleni Efthimiou, Stavroula-Evita Fotinea, Thomas Hanke, John Glauert, Richard Bowden, Annelies Braffort, Christophe Collet, Petros Maragos, and François Lefebvre-Albaret. 2012. The dicta-sign wiki: Enabling web communication for the deaf. InInternational Conference on Computer...

    2012

  7. [15]

    Oussama ElGhoul and Mohamed Jemni. 2011. WebSign: A system to make and interpret signs using 3D Avatars. InProceedings of the Second International Workshop on Sign Language Translation and A vatar Technology (SLTAT), Dundee, UK, Vol. 23

    2011

  8. [16]

    Tom Erez, Yuval Tassa, and Emanuel Todorov. 2015. Simulation tools for model- based robotics: Comparison of bullet, havok, mujoco, ode and physx. In2015 IEEE international conference on robotics and automation (ICRA). IEEE, 4397–4404

    2015

  9. [17]

    Steven Y Feng, Varun Gangal, Jason Wei, Sarath Chandar, Soroush Vosoughi, Teruko Mitamura, and Eduard Hovy. 2021. A survey of data augmentation approaches for NLP.arXiv preprint arXiv:2105.03075(2021)

    2021 arXiv

  10. [18]

    Aymane El Firdoussi, Mohamed El Amine Seddik, Soufiane Hayou, Reda Alami, Ahmed Alzubaidi, and Hakim Hacid. 2024. Maximizing the Potential of Synthetic Data: Insights from Random Matrix Theory.arXiv preprint arXiv:2410.08942 (2024)

    2024 arXiv

  11. [19]

    Edward Fish and Richard Bowden. 2025. Geo-Sign: Hyperbolic Con- trastive Regularisation for Geometrically Aware Sign Language Translation. arXiv:2506.00129 [cs.CV] https://arxiv.org/abs/2506.00129

    2025

  12. [20]

    Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. 2020. Generative adversarial networks.Commun. ACM63, 11 (2020), 139–144

    2020

  13. [21]

    Ian J Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. 2014. Generative adversarial nets.Advances in neural information processing systems27 (2014)

    2014

  14. [22]

    Jianping Gou, Baosheng Yu, Stephen J Maybank, and Dacheng Tao. 2021. Knowl- edge distillation: A survey.International Journal of Computer Vision129, 6 (2021), 1789–1819

    2021

  15. [23]

    Aaron Grattafiori, Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Alex Vaughan, et al. 2024. The llama 3 herd of models.arXiv preprint arXiv:2407.21783 (2024)

    2024 arXiv

  16. [24]

    Xi Guo, Wei Wu, Dongliang Wang, Jing Su, Haisheng Su, Weihao Gan, Jian Huang, and Qin Yang. 2022. Learning video representations of human motion from synthetic data. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 20197–20207

    2022

  17. [25]

    Aiming Hao, Yuecong Min, and Xilin Chen. 2021. Self-mutual distillation learn- ing for continuous sign language recognition. InProceedings of the IEEE/CVF international conference on computer vision. 11303–11312

    2021

  18. [26]

    Shuang Hao, Wenfeng Han, Tao Jiang, Yiping Li, Haonan Wu, Chunlin Zhong, Zhangjun Zhou, and He Tang. 2024. Synthetic data in AI: Challenges, applications, and ethical implications.arXiv preprint arXiv:2401.01629(2024)

    2024 arXiv

  19. [27]

    Geoffrey Hinton, Oriol Vinyals, and Jeff Dean. 2015. Distilling the knowledge in a neural network.arXiv preprint arXiv:1503.02531(2015)

    2015 arXiv

  20. [28]

    Jonathan Ho, Ajay Jain, and Pieter Abbeel. 2020. Denoising diffusion probabilistic models.Advances in neural information processing systems33 (2020), 6840–6851

    2020

  21. [29]

    Jonathan Ho, Tim Salimans, Alexey Gritsenko, William Chan, Mohammad Norouzi, and David J Fleet. 2022. Video diffusion models.Advances in Neu- ral Information Processing Systems35 (2022), 8633–8646

    2022

  22. [30]

    Hezhen Hu, Zhiwen Fan, Tianhao Wu, Yihan Xi, Seoyoung Lee, Georgios Pavlakos, and Zhangyang Wang. 2024. Expressive Gaussian Human Avatars from Monocular RGB Video. InNeurIPS

    2024

  23. [31]

    Lianyu Hu, Liqing Gao, Zekang Liu, and Wei Feng. 2022. Temporal lift pooling for continuous sign language recognition. InEuropean conference on computer vision. Springer, 511–527

    2022

  24. [32]

    Qisheng Hu, Kaixin Li, Xu Zhao, Yuxi Xie, Tiedong Liu, Hui Chen, Qizhe Xie, and Junxian He. 2023. Instructcoder: Empowering language models for code editing. CoRR(2023)

    2023

  25. [33]

    Wencan Huang, Wenwen Pan, Zhou Zhao, and Qi Tian. 2021. Towards Fast and High-Quality Sign Language Production. InProceedings of the 29th ACM International Conference on Multimedia. 3172–3181

    2021

  26. [34]

    Alfarabi Imashev, Medet Mukushev, Vadim Kimmelman, and Anara Sandygulova

  27. [35]

    Maksym Ivashechkin, Oscar Mendez, and Richard Bowden. 2023. Improving 3D Pose Estimation For Sign Language. In2023 IEEE International Conference on Acoustics, Speech, and Signal Processing Workshops (ICASSPW). 1–5. doi:10.1109/ ICASSPW59220.2023.10193629

    2023

  28. [36]

    Maksym Ivashechkin, Oscar Mendez, and Richard Bowden. 2025. SignSplat: Rendering Sign Language via Gaussian Splatting.arXiv preprint arXiv:2505.02108 (2025)

    2025 arXiv

  29. [37]

    Hamid Reza Vaezi Joze and Oscar Koller. 2018. Ms-asl: A large-scale data set and benchmark for understanding american sign language.arXiv preprint arXiv:1812.01053(2018)

    2018 arXiv

  30. [38]

    Bernhard Kerbl, Georgios Kopanas, Thomas Leimkühler, and George Drettakis

  31. [39]

    Yo-whan Kim, Samarth Mishra, SouYoung Jin, Rameswar Panda, Hilde Kuehne, Leonid Karlinsky, Venkatesh Saligrama, Kate Saenko, Aude Oliva, and Rogerio Feris. 2022. How transferable are video representations based on synthetic data? Advances in Neural Information Processing Syste...

    2022

  32. [40]

    Diederik P Kingma, Max Welling, et al. 2013. Auto-encoding variational bayes

    2013

  33. [41]

    Michael Kipp, Quan Nguyen, Alexis Heloir, and Silke Matthes. 2011. Assessing the deaf user perspective on sign language avatars. InThe proceedings of the 13th international ACM SIGACCESS conference on Computers and accessibility. 107–114

    2011

  34. [42]

    Oscar Koller, Necati Cihan Camgoz, Hermann Ney, and Richard Bowden. 2019. Weakly supervised learning with multi-stream CNN-LSTM-HMMs to discover sequential parallelism in sign language videos.IEEE transactions on pattern analysis and machine intelligence42, 9 (2019), 2306–2320

    2019

  35. [43]

    Reiner Konrad, Thomas Hanke, Gabriele Langer, Dolly Blanck, Julian Bleicken, Ilona Hofmann, Olga Jeziorski, Lutz König, Susanne König, Rie Nishio, Anja Regen, Uta Salden, Sven Wagner, Satu Worseck, Oliver Böse, Elena Jahn, and Marc Schulder. 2020. MEINE DGS – annotiert. Öffent...

    2020 doi

  36. [44]

    Dongxu Li, Cristian Rodriguez, Xin Yu, and Hongdong Li. 2020. Word-level deep sign language recognition from video: A new large-scale dataset and methods comparison. InProceedings of the IEEE/CVF winter conference on applications of computer vision. 1459–1469

    2020

  37. [45]

    Zecheng Li, Wengang Zhou, Weichao Zhao, Kepeng Wu, Hezhen Hu, and Houqiang Li. 2025. Uni-Sign: Toward Unified Sign Language Understanding at Scale. arXiv:2501.15187 [cs.CV] https://arxiv.org/abs/2501.15187 9 , , Harry Walsh, Maksym Ivashechkin, and Richard Bowden

    2025 arXiv

  38. [46]

    Chin-Yew Lin. 2004. Rouge: A package for automatic evaluation of summaries. InText summarization branches out. 74–81

    2004

  39. [47]

    Ruibo Liu, Jerry Wei, Fangyu Liu, Chenglei Si, Yanzhe Zhang, Jinmeng Rao, Steven Zheng, Daiyi Peng, Diyi Yang, Denny Zhou, et al. 2024. Best practices and lessons learned on synthetic data.arXiv preprint arXiv:2404.07503(2024)

    2024 arXiv

  40. [48]

    Yinhan Liu, Jiatao Gu, Naman Goyal, Xian Li, Sergey Edunov, Marjan Ghazvinine- jad, Mike Lewis, and Luke Zettlemoyer. 2020. Multilingual Denoising Pre-training for Neural Machine Translation.Transactions of the Association for Computational Linguistics8 (2020), 726–742. doi:10...

    2020 doi

  41. [49]

    Camillo Lugaresi, Jiuqiang Tang, Hadon Nash, Chris McClanahan, Esha Uboweja, Michael Hays, Fan Zhang, Chuo-Ling Chang, Ming Guang Yong, Juhyun Lee, et al. 2019. Mediapipe: A framework for building perception pipelines.arXiv preprint arXiv:1906.08172(2019)

    2019 arXiv

  42. [50]

    Haipeng Luo, Qingfeng Sun, Can Xu, Pu Zhao, Jianguang Lou, Chongyang Tao, Xiubo Geng, Qingwei Lin, Shifeng Chen, and Dongmei Zhang. 2023. Wizardmath: Empowering mathematical reasoning for large language models via reinforced evol-instruct.arXiv preprint arXiv:2308.09583(2023)

    2023 arXiv

  43. [51]

    Yuecong Min, Aiming Hao, Xiujuan Chai, and Xilin Chen. 2021. Visual align- ment constraint for continuous sign language recognition. Inproceedings of the IEEE/CVF international conference on computer vision. 11542–11551

    2021

  44. [52]

    Gyeongsik Moon, Takaaki Shiratori, and Shunsuke Saito. 2024. Expressive Whole- Body 3D Gaussian Avatar. InECCV

    2024

  45. [53]

    Amit Moryossef, Kayo Yin, Graham Neubig, and Yoav Goldberg. 2021. Data augmentation for sign language gloss translation.arXiv preprint arXiv:2105.07476 (2021)

    2021 arXiv

  46. [54]

    Michael Niemeyer and Andreas Geiger. 2021. Giraffe: Representing scenes as compositional generative neural feature fields. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition. 11453–11464

    2021

  47. [55]

    Kishore Papineni, Salim Roukos, Todd Ward, and Wei-Jing Zhu. 2002. Bleu: a method for automatic evaluation of machine translation. InProceedings of the 40th annual meeting of the Association for Computational Linguistics. 311–318

    2002

  48. [56]

    Georgios Pavlakos, Vasileios Choutas, Nima Ghorbani, Timo Bolkart, Ahmed A. A. Osman, Dimitrios Tzionas, and Michael J. Black. 2019. Expressive Body Capture: 3D Hands, Face, and Body from a Single Image. InProceedings IEEE Conf. on Computer Vision and Pattern Recognition (CVPR)

    2019

  49. [57]

    Georgios Pavlakos, Dandan Shan, Ilija Radosavovic, Angjoo Kanazawa, David Fouhey, and Jitendra Malik. 2024. Reconstructing Hands in 3D with Transformers. InCVPR

    2024

  50. [58]

    Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. 2021. Learning Transferable Visual Models From Natural Language Supervision.CoRRabs/2103.00...

    2021 arXiv

  51. [59]

    Javier Romero, Dimitrios Tzionas, and Michael J. Black. 2017. Embodied Hands: Modeling and Capturing Hands and Bodies Together.ACM Transactions on Graphics, (Proc. SIGGRAPH Asia)36, 6 (Nov. 2017)

    2017

  52. [60]

    Ben Saunders, Necati Cihan Camgöz, and Richard Bowden. 2020. Adversarial Training for Multi-Channel Sign Language Production. InBritish Machine Vision Virtual Conference

    2020

  53. [61]

    Ben Saunders, Necati Cihan Camgoz, and Richard Bowden. 2020. Everybody Sign Now: Translating Spoken Language to Photo Realistic Sign Language Video. arXiv preprint arXiv:2011.09846(2020)

    2020 arXiv

  54. [62]

    Ben Saunders, Necati Cihan Camgoz, and Richard Bowden. 2020. Progressive transformers for end-to-end sign language production. InComputer Vision–ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part XI 16. Springer, 687–705

    2020

  55. [63]

    Ben Saunders, Necati Cihan Camgoz, and Richard Bowden. 2021. Continuous 3d multi-channel sign language production via progressive transformers and mixture density networks.IJCV129, 7 (2021), 2113–2135

    2021

  56. [64]

    Ben Saunders, Necati Cihan Camgoz, and Richard Bowden. 2022. Signing at Scale: Learning to Co-Articulate Signs for Large-Scale Photo-Realistic Sign Language Production. InProceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR)

    2022

  57. [65]

    Ben Saunders, Necati Cihan Camgoz, and Richard Bowden. 2022. Signing at scale: Learning to co-articulate signs for large-scale photo-realistic sign language production. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 5141–5151

    2022

  58. [66]

    Mohamed El Amine Seddik, Suei-Wen Chen, Soufiane Hayou, Pierre Youssef, and Merouane Debbah. 2024. How bad is training on synthetic data? a statistical analysis of language model collapse.arXiv preprint arXiv:2404.05090(2024)

    2024 arXiv

  59. [67]

    Connor Shorten, Taghi M Khoshgoftaar, and Borko Furht. 2021. Text data aug- mentation for deep learning.Journal of big Data8, 1 (2021), 101

    2021

  60. [68]

    Ilia Shumailov, Zakhar Shumaylov, Yiren Zhao, Yarin Gal, Nicolas Papernot, and Ross Anderson. 2023. The curse of recursion: Training on generated data makes models forget.arXiv preprint arXiv:2305.17493(2023)

    2023 arXiv

  61. [69]

    William C Stokoe. 1980. Sign Language Structure.Annual Review of Anthropology (1980)

    1980

  62. [70]

    Stephanie Stoll, Necati Cihan Camgöz, Simon Hadfield, and R. Bowden. 2018. Sign Language Production using Neural Machine Translation and Generative Adver- sarial Networks. InBritish Machine Vision Conference. https://api.semanticscholar. org/CorpusID:52288950

    2018

  63. [71]

    Stephanie Stoll, Necati Cihan Camgoz, Simon Hadfield, and Richard Bowden

  64. [72]

    Shengeng Tang, Dan Guo, Richang Hong, and Meng Wang. 2021. Graph-Based Multimodal Sequential Embedding for Sign Language Translation.IEEE Transac- tions on MultimediaPP (10 2021), 1–1. doi:10.1109/TMM.2021.3117124

    2021

  65. [73]

    Shengeng Tang, Jiayi He, Lechao Cheng, Jingjing Wu, Dan Guo, and Richang Hong. 2024. Discrete to Continuous: Generating Smooth Transition Poses from Sign Language Observation.arXiv preprint arXiv:2411.16810(2024)

    2024 arXiv

  66. [74]

    Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Łukasz Kaiser, and Illia Polosukhin. 2017. Attention is all you need.Advances in neural information processing systems30 (2017)

    2017

  67. [75]

    Text2Sign: Towards Sign Language Production Using Neural Machine Translation and Generative Adversarial Networks.Int. J. Comput. Vision128, 4 (apr 2020), 891–908. doi:10.1007/s11263-019-01281-2

    2020 doi

  68. [76]

    Harry Walsh, Ben Saunders, and Richard Bowden. 2024. Sign Stitching: A Novel Approach to Sign Language Production. InThe 35th British Machine Vision Con- ference (BMVC)

    2024

  69. [77]

    Ryan Wong, Necati Cihan Camgoz, and Richard Bowden. 2024. Sign2GPT: Lever- aging large language models for gloss-free sign language translation.arXiv preprint arXiv:2405.04164(2024)

    2024 arXiv

  70. [78]

    Ryan Wong, Necati Cihan Camgoz, and Richard Bowden. 2025. SignRep: En- hancing Self-Supervised Sign Representations.arXiv preprint arXiv:2503.08529 (2025)

    2025 arXiv

  71. [79]

    Harry Walsh, Abolfazl Ravanshad, Mariam Rahmani, and Richard Bowden. 2024. A Data-Driven Representation for Sign Language Production. InProceedings of the 18th International Conference on Automatic Face and Gesture Recognition (FG 2024). Institute of Electrical and Electronics...

    2024

  72. [80]

    Huijie Yao, Wengang Zhou, Hao Zhou, and Houqiang Li. 2024. Semi-Supervised Spoken Language Glossification.arXiv preprint arXiv:2406.08173(2024)

    2024 arXiv

  73. [81]

    Longhui Yu, Weisen Jiang, Han Shi, Jincheng Yu, Zhengying Liu, Yu Zhang, James T Kwok, Zhenguo Li, Adrian Weller, and Weiyang Liu. 2023. Metamath: Bootstrap your own mathematical questions for large language models.arXiv preprint arXiv:2309.12284(2023)

    2023 arXiv

  74. [82]

    Jan Zelinka and Jakub Kanis. 2020. Neural Sign Language Synthesis: Words Are Our Glosses. In2020 IEEE Winter Conference on Applications of Computer Vision (W ACV). 3384–3392. doi:10.1109/WACV45572.2020.9093516

    2020

  75. [83]

    Mengde Xu, Zheng Zhang, Fangyun Wei, Yutong Lin, Yue Cao, Han Hu, and Xiang Bai. 2021. A simple baseline for zero-shot semantic segmentation with pre-trained vision-language model.arXiv preprint arXiv:2112.14757(2021)

    2021 arXiv

  76. [84]

    Inge Zwitserlood, Margriet Verlinden, Johan Ros, Sanny Van Der Schoot, and T Netherlands. 2004. Synthetic signing for the deaf: Esign. InProceedings of the conference and workshop on assistive technologies for vision and hearing impairment (CVHI). 10

    2004

  77. [87]

    Benjia Zhou, Zhigang Chen, Albert Clapés, Jun Wan, Yanyan Liang, Sergio Escalera, Zhen Lei, and Du Zhang. 2023. Gloss-free sign language translation: Improving from visual-language pretraining. InProceedings of the IEEE/CVF International Conference on Computer Vision. 20871–20881

    2023

  78. [2020]

    InProceedings of the 24th Conference on Computational Natural Language Learning

    K-RSL: a corpus for linguistic understanding, visual evaluation, and recog- nition of sign languages. InProceedings of the 24th Conference on Computational Natural Language Learning. Association for Computational Linguistics

  79. [2021]

    Bbc-oxford british sign language dataset.arXiv preprint arXiv:2111.03635 (2021)

    2021 arXiv

  80. [2023]

    https://repo-sam.inria.fr/fungraph/3d- gaussian-splatting/

    3D Gaussian Splatting for Real-Time Radiance Field Rendering.ACM Transactions on Graphics42, 4 (July 2023). https://repo-sam.inria.fr/fungraph/3d- gaussian-splatting/

    2023

Pith tools

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