Pith. sign in

REVIEW 3 major objections 4 minor 45 references

Contrastive Pretraining with Dual Visual Encoders for Gloss-Free Sign Language Translation

T0 review · 3 major / 4 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read This paper claims that fusing a spatial and a spatio-temporal visual encoder through contrastive alignment improves gloss-free sign language translation, reaching BLEU-4 23.81 on Phoenix-2014T.

desk verdict A solid internal ablation, but an unmatched I3D pretraining makes the gloss-free SOTA claim over-strong. read the letter →

arxiv 2507.10306 v1 pith:XDEKEG3U submitted 2025-07-14 cs.CV

Ozge Mercanoglu Sincan , Richard Bowden This is my paper
classification cs.CV
keywords SignLanguageTranslationGloss-freeSLTContrastiveLearningDualVisualEncodersMultimodalPretrainingPhoenix-2014TmBARTFeatureAggregation
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 is trying to establish that gloss-free sign language translation can be improved by pretraining two complementary visual encoders together with language, rather than relying on a single visual backbone or on gloss annotations. On the Phoenix-2014T benchmark, the proposed DVE-SLT system reports BLEU-4 23.81, the highest among the gloss-free methods it compares against, and consistently beats its own single-encoder variants. A sympathetic reader would care because glosses are expensive to annotate and simplify continuous signing, so a method that removes that supervision while improving translation suggests that semantic alignment, not gloss labels, is the load-bearing ingredient. The key comparison is between the fused dual-encoder model and either encoder alone, plus the contrast with prior gloss-free systems.

What carries the argument

The load-bearing mechanism is the dual-objective contrastive pretraining. A shared transformer encoder processes both streams' outputs after temporal compression; two symmetric InfoNCE cross-modal losses align ResNet and I3D embeddings to mBART sentence embeddings, and a third inter-modal InfoNCE loss aligns the two visual streams. During translation, the pretrained shared encoder and a 3-layer mBART decoder are reused, with visual features fused by channel-wise concatenation and passed through a VL adapter.

What would settle it

Train DVE-SLT with the I3D branch initialized from random weights, or from ImageNet/Kinetics rather than MeineDGS, and compare BLEU-4 on Phoenix-2014T. If the score falls to the single-stream level around 22.7 while the dual-encoder structure is unchanged, the central claim that the contrastive dual-encoder design drives the gain would fail; if it stays near 23.8, the claim survives.

Watch

Extended reading notes

Core claim

The central claim is that two visual backbones—a ResNet18 capturing fine-grained spatial information from individual frames and an I3D capturing spatio-temporal dynamics—are complementary, and that aligning both with sentence-level text embeddings plus with each other during contrastive pretraining yields better downstream translation than either backbone alone. On Phoenix-2014T, the dual-encoder model reaches BLEU-4 23.81, beating the best gloss-free competitor in the study; the authors attribute the gain to the inter-modal alignment pushing the two views of the same video closer while separating different videos, giving the translation decoder a semantically richer visual input.

Load-bearing premise

The load-bearing premise is that the BLEU-4 gain comes from the dual-encoder contrastive design rather than from extra pretraining data, since the I3D branch is initialized with weights from a German sign language corpus (MeineDGS) and none of the compared gloss-free baselines report using that same initialization.

Editorial extensions

If this is right

  • If the claim holds, gloss-free SLT no longer needs gloss labels to gain from visual-language pretraining; aligning two visual streams with text is enough to improve BLEU-4.
  • The ablation predicts that any single visual encoder, ResNet or I3D alone, will underperform the fused pair, so future gloss-free systems should consider complementary backbones.
  • Concatenation fusion being best suggests preserving full channel information from both streams is more important than learned attention or additive mixing.
  • OneCycleLR reaching comparable scores in fewer epochs suggests the pretraining plus fine-tuning pipeline can be trained more cheaply than cosine annealing.
  • The model outperforms heavier LLM-based baselines with a 3-layer mBART decoder, suggesting lightweight decoders can compete if visual features are well aligned.

Reading between the lines

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

  • In my reading, the reported BLEU-4 gap is provisional until the I3D initialization is matched: the paper uses MeineDGS pretraining for I3D, and none of the listed baselines report the same pretraining, so a controlled random-init ablation is needed to attribute the gain to the dual-encoder design.
  • A natural next experiment is to apply the same dual-encoder contrastive recipe to other sign language datasets to see whether the Phoenix-2014T ranking transfers.
  • The inter-modal loss is closely related to density-reduction goals; probing retrieval or frame-level feature density before and after pretraining would show whether the second encoder actually separates visually similar signs.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

3 major / 4 minor

Summary. The paper proposes DVE-SLT, a two-stage gloss-free sign language translation framework. In the pretraining stage, two visual encoders (ResNet18 for spatial features and I3D for spatio-temporal features) are projected into a shared space and aligned with sentence-level mBART text embeddings through symmetric InfoNCE losses, plus an inter-modal contrastive loss between the two visual branches. In the translation stage, the visual features are fused (via element-wise sum, concatenation, or cross-attention), passed through a visual-language adapter, and decoded by an mBART-based encoder-decoder. On Phoenix-2014T, the paper reports BLEU-4 of 23.81 for the concatenation-based dual-encoder model, which it claims is the highest among existing gloss-free approaches, outperforming the best reported gloss-free baseline (SignCL, 22.74) by 0.84 BLEU-4. Internal ablations show that the dual-encoder model outperforms either single encoder alone and that concatenation is the best fusion strategy.

Significance. If the reported result is robust, the paper makes a useful empirical contribution: it shows that a relatively simple dual visual encoder with contrastive alignment can improve gloss-free SLT on a standard benchmark, and it provides a clean internal ablation of encoder choice and fusion strategy. The paper is generally well written and the experiments are internally consistent: the ResNet-only, I3D-only, and fused models in Table 2 are monotonic in BLEU-4, and Table 3 shows all fusion variants outperform the single encoders. However, the headline claim of 'highest among gloss-free approaches' is weakened by a potential pretraining confound that the paper discloses but does not control for, and the absence of variance estimates leaves the small reported margins difficult to interpret. The work would be strengthened by ablations that isolate the MeineDGS initialization and the two contrastive losses.

major comments (3)
  1. [Section 4.2 and Table 1] The comparison with prior gloss-free methods is confounded by the MeineDGS pretraining of the I3D encoder. Section 4.2 states: 'We initialized our I3D model with weights pre-trained on the MeineDGS German sign language dataset [15] for isolated sign language translation [31].' None of the baselines in Table 1 (GFSLT-VLP, SignCL, Sign2GPT, FLa-LLM, SignLLM) are reported as using MeineDGS or another large German Sign Language pretraining corpus. Since Phoenix-2014T is also German Sign Language, the additional pretraining data could plausibly account for the +0.84 BLEU-4 margin over SignCL. This is load-bearing because the central claim is based entirely on the cross-paper comparison. The authors should either reimplement or retrain a baseline with the same MeineDGS-initialized I3D, or add a DVE-SLT variant trained without MeineDGS initialization; without such a control, the 'highest among gloss-free' claim is not causally attributable to the dual-encoder design.
  2. [Table 2 and Section 3.1] The internal ablation also inherits the MeineDGS confound and does not isolate the contribution of the contrastive objectives. In Table 2, the I3D-only and ResNet+I3D rows use the MeineDGS-pretrained I3D, while the ResNet-only row does not, so the observed fusion gain over ResNet could be due to the pretrained I3D backbone rather than to dual-encoder fusion or inter-modal alignment. Additionally, the paper does not ablate the losses in Eq. (3): there is no variant without L_inter, no variant without L_cross, and no variant without any contrastive pretraining. The claim that 'dual-objective contrastive alignment' improves translation therefore is not directly tested. The authors should add ablations with randomly initialized I3D and with each loss term removed to support the architectural attribution.
  3. [Tables 1 and 4] The reported margins are small and no variance estimates are provided. The headline improvement over SignCL is +0.84 BLEU-4, and the best scheduler in Table 4 is selected from a sweep of several configurations (CosAnLR, ExpLR, OneCycleLR) with different batch sizes and learning rates. Cross-paper BLEU differences below one point are typically within run-to-run variation for this benchmark, and the final configuration is the outcome of a model-selection procedure. The authors should report multiple seeds (with mean and standard deviation) or at least state the number of runs and confirm that the reported improvements are stable.
minor comments (4)
  1. [Table 4] The table header contains a typo: 'OneCyleLR' should be 'OneCycleLR'.
  2. [Section 3.1] The caption of Figure 1 refers to 'shared visual encoder, E_shared' and 'Text Decoder weights' being utilized in the downstream task, but the text in Section 3.2 says the encoder and decoder are initialized with 'the pretrained shared encoder and text decoder.' It would help to clarify explicitly which components of the pretrained model (shared encoder, text decoder, adapters, temporal encoders) are transferred and which are re-initialized in the translation stage.
  3. [Section 2.2] The phrase 'Initial gloss-free efforts (Sign2Text)' is ambiguous; either expand the reference or clarify whether 'Sign2Text' is a specific method or a generic category.
  4. [Figure 2] The qualitative attention analysis is informative, but the y-axis label 'video frames' would be clearer if accompanied by a description of how frames are sampled or temporally aligned to the I3D windows, especially since the dual encoders operate at different temporal resolutions.

Circularity Check

0 steps flagged · score 2.0 of 10

No circular derivation; DVE-SLT's BLEU score is an empirical downstream result, and the only self-citation (MeineDGS I3D initialization) is a comparability confound, not an input-output identity.

full rationale

The paper's claimed derivation is empirical: contrastive losses in Eqs. (1)-(3) define a pretraining objective, and the headline BLEU-4=23.81 is computed on the Phoenix-2014T test set after fine-tuning. That metric is not one of the pretraining inputs, nor is it a parameter fitted by the loss, so the central claim is not self-definitional. Table 2 provides a controlled single-encoder versus dual-encoder comparison: the I3D-only row (22.71) and the ResNet+I3D row (23.81) share the same MeineDGS-initialized I3D backbone, so the dual-encoder gain is not an artifact of that initialization. The comparison to prior gloss-free methods in Table 1 is cross-paper and may be confounded by the authors' MeineDGS I3D pretraining [31], but a confound is a threat to external validity, not circularity: there is no equation in which the reported BLEU reduces to the pretraining choice by construction. The acknowledged weather-domain limitation in the ethical statement further concerns generalizability rather than the derivation. Apart from self-citations [30,31] as sources of backbones and pretraining, no load-bearing argument is justified solely by a self-citation, and no uniqueness theorem or ansatz is imported from the authors' prior work. Accordingly, there are no specific circular steps to list; the low score reflects only the presence of minor self-citations in the experimental setup.

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

The central claim depends more on experimental setup than on derived mathematics. The main unstated inputs are the pretraining data source (presumably Phoenix-2014T train), the MeineDGS I3D initialization, the mBART language model, and the selected scheduler hyperparameters. These are domain assumptions and fitting choices, not invented entities.

free parameters (4)
  • Trainable temperature tau = not reported
    Used as the temperature in both InfoNCE losses (Eq. 1 and Eq. 2); it is learned during pretraining and controls the sharpness of the alignment.
  • OneCycleLR pct_start = 0.35
    Selected after a scheduler sweep in Table 4; the headline BLEU-4 of 23.81 depends on this configuration, while CosAnLR gives 23.18.
  • Batch size = 8
    Used in pretraining and in the best translation run; batch size affects the number of negatives in the contrastive losses.
  • I3D sliding window stride = 6 frames
    Determines temporal overlap and how many I3D features are extracted per video; not ablated.
assumptions (4)
  • domain assumption Phoenix-2014T training split is the pretraining corpus
    The paper never explicitly states which dataset is used for contrastive pretraining, but the downstream evaluation is on Phoenix-2014T and pretraining and fine-tuning appear to share the same training split.
  • domain assumption I3D initialized on MeineDGS is a fair and helpful starting point
    Section 4.2 uses MeineDGS-pretrained I3D weights; the comparison assumes this external data does not break parity with baselines.
  • domain assumption mBART text encoder and decoder provide suitable language representations
    The pretrained multilingual mBART model is used as text encoder and decoder initialization; the method assumes its embeddings align well with visual features for German Sign Language text.
  • standard math Symmetric InfoNCE losses are appropriate objectives
    The contrastive losses (Eq. 1 and 2) assume in-batch negatives and a trainable temperature are sufficient to learn the alignment.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Contrastive Pretraining with Dual Visual Encoders for Gloss-Free Sign Language Translation." pith.science (2026). https://pith.science/paper/XDEKEG3U

@misc{pith2026250710306,
  author       = {Pith},
  title        = {Pith review of: Contrastive Pretraining with Dual Visual Encoders for Gloss-Free Sign Language Translation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/XDEKEG3U}},
  note         = {Machine review of arXiv:2507.10306}
}
read the original abstract

Sign Language Translation (SLT) aims to convert sign language videos into spoken or written text. While early systems relied on gloss annotations as an intermediate supervision, such annotations are costly to obtain and often fail to capture the full complexity of continuous signing. In this work, we propose a two-phase, dual visual encoder framework for gloss-free SLT, leveraging contrastive visual-language pretraining. During pretraining, our approach employs two complementary visual backbones whose outputs are jointly aligned with each other and with sentence-level text embeddings via a contrastive objective. During the downstream SLT task, we fuse the visual features and input them into an encoder-decoder model. On the Phoenix-2014T benchmark, our dual encoder architecture consistently outperforms its single stream variants and achieves the highest BLEU-4 score among existing gloss-free SLT approaches.

Figures

Figures reproduced from arXiv: 2507.10306 by the authors.

Figure 1
Figure 1. Overview of the proposed framework which has two phases: (a) pretraining, (b) translation. In pretraining, visual and [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Cross-attention map between video frames (y-axis) [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

45 extracted references · 33 canonical work pages

  1. [31]

    Ozge Mercanoglu Sincan, Necati Cihan Camgoz, and Richard Bowden. 2024. Using an LLM to Turn Sign Spottings into Spoken Language Sentences. arXiv preprint arXiv:2403.10434 (2024)

    work page Pith review arXiv 2024

  2. [15]

    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. Öffentliches Korpus der Deutschen Gebärdensprache, 3. Release / MY DGS – annotated. Pu...

    doi:10.25592/dgs.corpus-3.0 2020

  3. [1]

    World Federation of the Deaf

    2025. World Federation of the Deaf. https://wfdeaf.org/contact/faqs/ Accessed: 2025-05-30

    work page 2025

  4. [2]

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

  5. [3]

    Necati Cihan Camgoz, Simon Hadfield, Oscar Koller, Hermann Ney, and Richard Bowden. 2018. Neural Sign Language Translation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR)

    work page 2018

  6. [4]

    Necati Cihan Camgoz, Oscar Koller, Simon Hadfield, and Richard Bowden. 2020. Sign language transformers: Joint end-to-end sign language recognition and translation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 10023–10033

    work page 2020

  7. [5]

    Joao Carreira and Andrew Zisserman. 2017. Quo vadis, action recognition? a new model and the kinetics dataset. In proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) . 6299–6308

    work page 2017

  8. [6]

    Yutong Chen, Fangyun Wei, Xiao Sun, Zhirong Wu, and Stephen Lin. 2022. A simple multi-modality transfer learning baseline for sign language translation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition . 5120–5130

    work page 2022

Show all 45 references
  1. [7]

    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 Systems 35 (2022), 17043–17056

    2022

  2. [8]

    Zhigang Chen, Benjia Zhou, Jun Li, Jun Wan, Zhen Lei, Ning Jiang, Quan Lu, and Guoqing Zhao. 2024. Factorized Learning Assisted with Large Language Model for Gloss-free Sign Language Translation. the 2024 Joint International Conference on Computational Linguistics, Language Re...

    2024

  3. [9]

    Yiting Cheng, Fangyun Wei, Jianmin Bao, Dong Chen, and Wenqiang Zhang

  4. [10]

    Rohit Girdhar, Alaaeldin El-Nouby, Zhuang Liu, Mannat Singh, Kalyan Vasudev Alwala, Armand Joulin, and Ishan Misra. 2023. Imagebind: One embedding space to bind them all. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 15180–15190

    2023

  5. [11]

    Jia Gong, Lin Geng Foo, Yixuan He, Hossein Rahmani, and Jun Liu. 2024. Llms are good sign language translators. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) . 18362–18372

    2024

  6. [12]

    Michael Gutmann and Aapo Hyvärinen. 2010. Noise-contrastive estimation: A new estimation principle for unnormalized statistical models. In Proceedings of the thirteenth international conference on artificial intelligence and statistics . JMLR Workshop and Conference Proceeding...

    2010

  7. [13]

    Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. 2016. Deep residual learning for image recognition. InProceedings of the IEEE Conference on Computer Vision and Pattern Recognition . 770–778

    2016

  8. [14]

    Peiqi Jiao, Yuecong Min, and Xilin Chen. 2025. Visual Alignment Pre-training for Sign Language Translation. In European Conference on Computer Vision . Springer, 349–367

    2025

  9. [16]

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

    2004

  10. [17]

    Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. 2023. Visual in- struction tuning. Advances in neural information processing systems 36 (2023), 34892–34916

    2023

  11. [18]

    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 Linguistics 8 (2020), 726–742. http...

    2020 doi

  12. [19]

    Ilya Loshchilov and Frank Hutter. 2017. SGDR: Stochastic Gradient Descent with Warm Restarts. In International Conference on Learning Representations . https://openreview.net/forum?id=Skq89Scxx

    2017

  13. [20]

    Liliane Momeni, Hannah Bull, KR Prajwal, Samuel Albanie, Gül Varol, and An- drew Zisserman. 2022. Automatic dense annotation of large-vocabulary sign language videos. In European Conference on Computer Vision . Springer, 671–690

    2022

  14. [21]

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

    2002

  15. [22]

    Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. 2019. Pytorch: An imperative style, high-performance deep learning library. Advances in neural information processing system...

    2019

  16. [23]

    Matt Post. 2018. A Call for Clarity in Reporting BLEU Scores. In Proceedings of the Third Conference on Machine Translation: Research Papers . Association for Computational Linguistics, Belgium, Brussels, 186–191. https://www.aclweb. org/anthology/W18-6319

    2018

  17. [24]

    Amy Pu, Hyung Won Chung, Ankur P Parikh, Sebastian Gehrmann, and Thibault Sellam. 2021. Learning compact metrics for MT. In Proceedings of EMNLP

    2021

  18. [25]

    Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. 2021. Learning transferable visual models from natural language supervision. In International Conference on Machine Learni...

    2021

  19. [26]

    Ludan Ruan, Anwen Hu, Yuqing Song, Liang Zhang, Sipeng Zheng, and Qin Jin

  20. [27]

    Thibault Sellam, Dipanjan Das, and Ankur P Parikh. 2020. BLEURT: Learning robust metrics for text generation. arXiv preprint arXiv:2004.04696 (2020)

    2020 arXiv

  21. [28]

    Bowen Shi, Diane Brentari, Gregory Shakhnarovich, and Karen Livescu. 2022. Open-Domain Sign Language Translation Learned from Online Video. In Pro- ceedings of the 2022 Conference on Empirical Methods in Natural Language Pro- cessing, Yoav Goldberg, Zornitsa Kozareva, and Yue ...

    2022 doi

  22. [29]

    In Proceedings of the AAAI Conference on Artificial Intelligence , Vol

    Accommodating audio modality in CLIP for multimodal processing. In Proceedings of the AAAI Conference on Artificial Intelligence , Vol. 37. 9641–9649

  23. [30]

    Ozge Mercanoglu Sincan, Necati Cihan Camgoz, and Richard Bowden. 2023. Is context all you need? scaling neural sign language translation to large domains of discourse. In Proceedings of the IEEE/CVF International Conference on Computer Vision. 1955–1965

    2023

  24. [32]

    Bowen Shi, Diane Brentari, Gregory Shakhnarovich, and Karen Livescu. 2022. Ttic’s wmt-slt 22 sign language translation system. In Proceedings of the Seventh Conference on Machine Translation (WMT) . 989–993

    2022

  25. [33]

    A Vaswani. 2017. Attention is all you need. Advances in Neural Information Processing Systems (2017)

    2017

  26. [34]

    Ryan Wong, Necati Cihan Camgoz, and Richard Bowden. 2024. Sign2GPT: Leveraging Large Language Models for Gloss-Free Sign Language Translation. In The Twelfth International Conference on Learning Representations . https: //openreview.net/forum?id=LqaEEs3UxU

    2024

  27. [35]

    Leslie N Smith and Nicholay Topin. 2019. Super-convergence: Very fast training of neural networks using large learning rates. InArtificial intelligence and machine learning for multi-domain operations applications , Vol. 11006. SPIE, 369–386

    2019

  28. [36]

    Huijie Yao, Wengang Zhou, Hao Feng, Hezhen Hu, Hao Zhou, and Houqiang Li

  29. [37]

    Jinhui Ye, Xing Wang, Wenxiang Jiao, Junwei Liang, and Hui Xiong. 2024. Improv- ing Gloss-free Sign Language Translation by Reducing Representation Density. In The Thirty-eighth Annual Conference on Neural Information Processing Systems . https://openreview.net/forum?id=FtzLbGoHW2

    2024

  30. [38]

    Hu Xu, Gargi Ghosh, Po-Yao Huang, Dmytro Okhonko, Armen Aghajanyan, Florian Metze, Luke Zettlemoyer, and Christoph Feichtenhofer. 2021. VideoCLIP: Contrastive Pre-training for Zero-shot Video-Text Understanding. In Proceedings of the 2021 Conference on Empirical Methods in Nat...

    2021

  31. [39]

    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. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) ....

    2023

  32. [40]

    In Proceedings of the IEEE/CVF International Conference on Computer Vision

    Sign language translation with iterative prototype. In Proceedings of the IEEE/CVF International Conference on Computer Vision . 15592–15601

  33. [41]

    Bin Zhu, Bin Lin, Munan Ning, Yang Yan, Jiaxi Cui, WANG HongFa, Yatian Pang, Wenhao Jiang, Junwu Zhang, Zongwei Li, Cai Wan Zhang, Zhifeng Li, Wei Liu, and Li Yuan. 2024. LanguageBind: Extending Video-Language Pretraining to N-modality by Language-based Semantic Alignment. In ...

    2024

  34. [42]

    Biao Zhang, Mathias Müller, and Rico Sennrich. 2022. SLTUNET: A Simple Unified Model for Sign Language Translation. In The Eleventh International Conference on Learning Representations (ICLR)

    2022

  35. [44]

    Hao Zhou, Wengang Zhou, Weizhen Qi, Junfu Pu, and Houqiang Li. 2021. Improv- ing Sign Language Translation With Monolingual Data by Sign Back-Translation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recog- nition. 1316–1325

    2021

  36. [2021]

    arXiv preprint arXiv:2111.03635 (2021)

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

    2021 arXiv

  37. [2023]

    In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition

    Cico: Domain-aware sign language retrieval via cross-lingual contrastive learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 19016–19026

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.