Pith. sign in

REVIEW 3 major objections 5 minor 247 references

Sign-language QA: answering natural-language questions about sign videos tests understanding beyond fixed gloss and translation mappings.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · deepseek-v4-flash

2026-08-01 00:42 UTC pith:VDIJOP4E

load-bearing objection Useful SLQA benchmark and baseline, but the 'beyond recognition/translation' claim needs a text-only oracle experiment to hold up. the 3 major comments →

arxiv 2607.27826 v1 pith:VDIJOP4E submitted 2026-07-30 cs.AI cs.CV

Sign Language Question Answering: A New Task, Benchmark, and Baseline for Sign Language Understanding

Shiwei Gan , Lichen Wang , Xiao Liu , Yafeng Yin , Kuizhuang Liu , Sanglu Lu , Lei Xie This is my paper
classification cs.AI cs.CV
keywords sign language question answeringsign language understandingvideo question answeringtemporal reasoningbenchmark constructionknowledge transferQCMTDcontinuous sign language recognition
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

This paper introduces Sign Language Question Answering (SLQA), a new evaluation task in which a model must answer natural-language questions about a sign-language video. The authors construct two benchmark datasets by automatically converting existing gloss and translation annotations into five complementary question types, then propose a baseline that pre-trains on sign recognition and translation before fine-tuning on QA. Their central claim is that question answering provides a more flexible and comprehensive test of sign-language understanding than either recognition or translation alone. The results show that the proposed baseline consistently outperforms general-purpose video-language models across all question categories, establishing a strong starting point for this new task. A sympathetic reader would care because this reframes sign-language understanding as a queryable and interactive capability rather than a one-shot mapping to text.

Core claim

The paper's central discovery is that sign-language understanding can be evaluated through question answering: given a sign-language video and an arbitrary natural-language question, a model should generate the correct answer. To support this, the authors produce SignQA, a benchmark of 144,550 template-generated question-answer pairs covering position reasoning, structural reasoning, visual search, gloss recognition, and translation understanding. Their baseline, SLQAM, uses a Question-Conditioned Modulated Temporal Downsampling (QCMTD) module and a three-stage in-domain knowledge transfer pipeline, and it outperforms representative video-language models and a cascaded sign-to-text-to-answer

What carries the argument

The central mechanism is the Question-Conditioned Modulated Temporal Downsampling (QCMTD) module. Before each temporal pooling layer, frame features are modulated by a global question embedding and per-frame relevance weights, producing a question-aware video representation that preserves segments relevant to the question. This replaces the fixed, question-agnostic temporal convolution used in prior sign-language pipelines. The other key machinery is a three-stage in-domain knowledge transfer strategy: first a visual encoder is pre-trained with continuous sign-language recognition using CTC loss, then a language decoder is trained for sign-language translation, and finally the whole model is

Load-bearing premise

The load-bearing premise is that question-answer pairs generated from existing gloss and translation annotations actually measure video-grounded semantic understanding rather than, for many categories, just recognition and text-based lookup.

What would settle it

Run an oracle experiment in which the ground-truth gloss sequence and translation are provided as text to the answer generator, skipping the video entirely. If a text-only oracle achieves near-perfect accuracy on M1, M2, and M3, those categories would be shown to test recognition-based lookup rather than video-grounded reasoning.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • SLQA can serve as a standardized benchmark for evaluating sign-language understanding beyond fixed recognition and translation objectives.
  • Question-conditioned temporal downsampling is shown to be more effective than fixed temporal convolution or Q-Former-style aggregation for sign-video question answering.
  • In-domain pre-training on continuous sign-language recognition and translation provides a substantial boost over training from scratch or from generic video-language models.
  • The five question categories offer a diagnostic breakdown of model strengths, from fine-grained visual search and sequential reasoning to sentence-level semantic understanding.
  • The token-weighted fine-tuning strategy is a practical recipe for training on template-generated QA benchmarks, preventing models from memorizing answer patterns.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • Because M1-M3 questions are deterministic lookups over gloss sequences, a model with strong gloss recognition and a small text-reasoning module could solve them without true video-grounded understanding; an oracle experiment feeding ground-truth glosses as text would reveal how much of the benchmark is actually testing recognition versus reasoning.
  • The template-based construction limits linguistic and visual diversity; extending SignQA with human-written open-ended questions and free-form answers would strengthen claims about general sign-language understanding.
  • The QCMTD mechanism could transfer to other fine-grained video-QA tasks where questions target short temporal segments, such as gesture recognition or sports-event reasoning.
  • The per-category results suggest that future SLU models should be compared on their reasoning profiles across these five capabilities, not just on translation quality.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 5 minor

Summary. The paper introduces Sign Language Question Answering (SLQA), a new task in which a model must answer natural-language questions about a sign-language video. The authors construct two benchmarks, SignQA-PHOENIX14T and SignQA-CSL-Daily, by instantiating 30 templates over existing gloss and sentence annotations, covering position reasoning, structural reasoning, visual search, gloss recognition, and translation understanding. They also propose a baseline, SLQAM, with a question-conditioned temporal downsampling module (QCMTD) and a three-stage in-domain knowledge transfer pipeline (CSLR pretraining, SLT training, SLQA fine-tuning). Experiments compare SLQAM with general-purpose VLMs and a cascaded Sign2Text2Answer baseline, reporting consistent improvements across categories and overall metrics.

Significance. If the benchmark truly measures video-grounded semantic understanding beyond recognition and translation, this is a valuable contribution: it defines a new evaluation paradigm, provides two public benchmarks, and gives a strong baseline. The paper also has practical strengths: the datasets are released, random sampling uses fixed seeds, and the experimental setup is clearly described. However, the central claim is not yet established. Because every QA pair is derived from the same gloss and translation annotations that define recognition and translation, the benchmark may be solvable by a system that first recognizes glosses/translations and then performs text-only QA; the paper reports no oracle experiment to rule this out. In addition, the reported performance differences are not accompanied by error bars or significance tests. These are load-bearing gaps for the paper's headline claims.

major comments (3)
  1. [Dataset Construction, Table 10] The benchmark's construct validity is unverified. Each of the five question categories is generated from the existing gloss or sentence annotations: M4 asks for the gloss sequence itself, M5 asks for the spoken translation itself, and M1–M3 are deterministic lookups over the gloss sequence (e.g., 'Which gloss appears at position 2?'). Thus an oracle that is given ground-truth glosses and translations as text-only input to the answer generator would answer the benchmark without any video understanding. The paper claims SLQA assesses 'multiple reasoning capabilities beyond recognition and translation' (Abstract, Introduction), but no such oracle upper bound is reported. The authors should add this decisive control, and also analyze which categories cannot be solved by the oracle, to substantiate the central claim.
  2. [Tables 5–8] The 'consistently outperforms' claim lacks statistical support. All tables report a single run with no error bars, confidence intervals, or significance tests. Some differences are small—for example, on PHOENIX14T test overall, SLQAM achieves BLEU-4 34.12 vs. 33.24 for Sign2Text2Answer (Table 6), and on CSL-Daily test overall, ROUGE-L is 73.52 vs. 70.75 (Table 8). Given that the benchmark is template-generated, paired bootstrap or multiple-seed reporting is essential to show the gains are not noise. This is a load-bearing issue for the main empirical comparison.
  3. [In-domain Knowledge Transfer Training / Dataset Construction] There is a potential train/evaluation overlap that weakens the interpretation of the baseline results. The model is pretrained on CSLR and SLT using PHOENIX14T and CSL-Daily annotations, and then fine-tuned on QA pairs generated from the same annotations. Since M4 and M5 ask for exactly those annotations, strong performance may reflect memorized recognition/translation mappings rather than question-conditioned video reasoning. A text-only oracle on the QA head, or evaluation on a held-out set with questions that cannot be derived from the original annotations, would clarify whether the QCMTD and pretraining gains are due to genuine question-aware video understanding or to annotation reuse.
minor comments (5)
  1. [Tables 5–6 captions] 'Phenix14T' should be 'Phoenix14T'.
  2. [Evaluation of SLQA Performance on CSL-Daily-QA] The sentence 'Tables 7 and 8 compare SLQAM with general-purpose VLMs and the cascaded Sign2Text2Answer baseline on CSL-Daily-QA da.' is incomplete; 'da' should be 'dataset'.
  3. [Table 9] The 'Total QA Pairs' row reads '3035,480'; this likely means '30 templates; 35,480 pairs' but the formatting is ambiguous and should be corrected.
  4. [Related Work, paragraph 1] The phrase 'we formulate SLQA as a bora SLU task' contains a typo; presumably 'novel' was intended.
  5. [Dataset Construction, M3] The non-manual marker questions on CSL-Daily are said to be instantiated from annotations, but the procedure for deriving the interrogative/non-interrogative answer is not specified. Please clarify how the template determines the answer from the original annotation without additional manual labeling.

Circularity Check

1 steps flagged

The 'beyond recognition and translation' claim is partially definitional: M4/M5 are the recognition/translation tasks restated as questions, and M1–M3 are deterministic lookups over the same gloss annotations.

specific steps
  1. self definitional [Dataset Construction (Question Categories M4/M5; Generation Procedure); Abstract]
    "M4 – Gloss Recognition. The model answers questions that require recognizing the gloss sequence corresponding to the input SL video. ... M5 – Translation Understanding. The model answers questions requiring the spoken-language translation of the input SL video. ... The answer is then deterministically derived from the annotations according to the question type."

    SLQA is introduced as evaluating SLU 'beyond recognition and translation,' but M4's correct answer is by construction the GT gloss sequence (CSLR output) and M5's correct answer is the GT spoken sentence (SLT output). M1–M3 are likewise deterministic operations on the same GT gloss sequence (fetch-by-position, neighbor, occurrence search). Hence the benchmark's advertised 'multiple reasoning capabilities beyond recognition and translation' reduces, for all five categories, to recognizing/translating the input annotations or performing string lookups on them; the QA wrapper does not add an independently specified reasoning target.

full rationale

There is no formal derivation chain in this paper: the contribution is a task formulation, automatically generated benchmarks, and an empirical baseline. The baseline experiments are internally coherent and are compared against externally fine-tuned VLMs (VideoLLaMA3, Qwen3-VL, InternVL3), so the engineering and evaluation sections are not circular. The circularity/construct-validity issue is confined to the central benchmarking claim. The paper states that QA pairs are generated 'by leveraging the existing gloss and sentence annotations' and that M4 and M5 ask for the gloss sequence and the spoken-language translation. Those are the very recognition/translation outputs the paper says SLQA goes beyond, and the remaining categories are deterministic queries over the same gloss annotation. Thus the assertion that SLQA 'assesses multiple reasoning capabilities beyond recognition and translation' is not demonstrated; a text-only oracle using ground-truth glosses/translations would be the decisive control and is absent. Weighing this against the substantial independent benchmark/baseline work, the score reflects a partial, definitional reduction rather than a full equivalence or a self-citation chain.

Axiom & Free-Parameter Ledger

2 free parameters · 4 axioms · 0 invented entities

The benchmark’s validity rests on two domain assumptions: that the source annotations are correct and that template-instantiated questions from them actually probe video-grounded understanding. The paper does not test the second assumption with a gold-text oracle. The remaining axioms are standard ML tooling assumptions (CTC, pretrained models). No invented physical entities are introduced.

free parameters (2)
  • Template/answer token loss weights = 0.02 (template), 0.98 (answer)
    Stage 3 loss weighting (Eq. 3) chosen by hand; it downweights template tokens and encodes test-time template structure into training.
  • Augmentation hyperparameters = flip p=0.5; temporal scaling ±20%
    Data preprocessing choices; horizontal flip is linguistically questionable for sign language because it mirrors handedness.
axioms (4)
  • domain assumption Ground-truth gloss and sentence annotations in PHOENIX14T and CSL-Daily are correct and sufficient to generate unambiguous QA pairs.
    Dataset Construction relies entirely on these annotations; no manual verification is described.
  • domain assumption Template-instantiated questions are a valid probe of video-grounded sign language understanding.
    Question Categories claims this, but M4/M5 are annotation restatements and no oracle baseline tests the assumption.
  • standard math CTC loss aligns video frames to glosses for pretraining.
    Eq. (1); standard assumption for CSLR.
  • domain assumption Pre-trained ResNet-18 and mT5 provide appropriate initialization for the sign-language QA model.
    Implementation Details; no evidence they are optimal for sign-language features.

pith-pipeline@v1.3.0-daily-deepseek · 19257 in / 14266 out tokens · 123790 ms · 2026-08-01T00:42:09.020724+00:00 · methodology

0 comments
read the original abstract

Recent advances in sign language (SL) understanding (SLU) have led to remarkable progress in tasks such as continuous SL recognition and SL translation. However, these tasks are designed with predefined objectives, requiring models to learn a fixed mapping from sign videos to glosses or spoken-language sentences. As a result, they provide only a limited assessment of whether a model truly understands the semantic content of SL videos. To address this limitation, \textbf{we first propose a new task, Sign Language Question Answering (SLQA)}, which evaluates SL understanding by requiring models to answer arbitrary natural language questions about SL videos. Unlike previous SLU tasks, SLQA provides a more flexible and comprehensive evaluation framework that assesses multiple reasoning capabilities beyond recognition and translation. To facilitate this task, \textbf{we further construct two SignQA benchmarks} based on PHOENIX14T and CSL-Daily by automatically generating question-answer pairs from existing gloss and sentence annotations using carefully designed templates. The resulting datasets cover five complementary question categories, including position reasoning, structural reasoning, visual search, gloss recognition, and translation understanding. \textbf{Finally, we propose a simple yet effective baseline model} equipped with a Question-Conditioned Modulated Temporal Downsampling module and an in-domain knowledge transfer strategy, enabling effective knowledge transfer from existing SLU tasks while enhancing question-aware temporal feature modeling. Extensive experiments demonstrate that our baseline consistently outperforms representative vision-language models across all question categories, establishing a strong benchmark for future research on SLQA. Datasets are available at:{https://huggingface.co/datasets/hulala/SignQA-2026}.

Figures

Figures reproduced from arXiv: 2607.27826 by Kuizhuang Liu, Lei Xie, Lichen Wang, Sanglu Lu, Shiwei Gan, Xiao Liu, Yafeng Yin.

Figure 1
Figure 1. Figure 1: Illustration of existing SLU and proposed SLQA. [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Overall architecture of the proposed SLQAM. The framework consists of a QCMTD module and an in-domain [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Illustration of each question category in SLQA. [PITH_FULL_IMAGE:figures/full_fig_p011_3.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

247 extracted references · 34 linked inside Pith

  1. [1]

    Proceedings of the 2022 conference on empirical methods in natural language processing , pages=

    Video question answering: Datasets, algorithms and challenges , author=. Proceedings of the 2022 conference on empirical methods in natural language processing , pages=

    2022

  2. [2]

    Proceedings of the 31st International Conference on Computational Linguistics , pages=

    VideoQA-TA: Temporal-Aware Multi-Modal Video Question Answering , author=. Proceedings of the 31st International Conference on Computational Linguistics , pages=

  3. [3]

    arXiv preprint arXiv:2404.03413 , year=

    MiniGPT4-Video: Advancing Multimodal LLMs for Video Understanding with Interleaved Visual-Textual Tokens , author=. arXiv preprint arXiv:2404.03413 , year=

    Pith/arXiv arXiv

  4. [4]

    European Conference on Computer Vision , pages=

    Llama-vid: An image is worth 2 tokens in large language models , author=. European Conference on Computer Vision , pages=. 2024 , organization=

    2024

  5. [5]

    Proceedings of the IEEE international conference on computer vision , pages=

    Vqa: Visual question answering , author=. Proceedings of the IEEE international conference on computer vision , pages=

  6. [6]

    Proceedings of the IEEE conference on computer vision and pattern recognition , pages=

    Clevr: A diagnostic dataset for compositional language and elementary visual reasoning , author=. Proceedings of the IEEE conference on computer vision and pattern recognition , pages=

  7. [7]

    Advances in neural information processing systems , volume=

    Vilbert: Pretraining task-agnostic visiolinguistic representations for vision-and-language tasks , author=. Advances in neural information processing systems , volume=

  8. [8]

    International conference on machine learning , pages=

    Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation , author=. International conference on machine learning , pages=. 2022 , organization=

    2022

  9. [9]

    Advances in neural information processing systems , volume=

    Flamingo: a visual language model for few-shot learning , author=. Advances in neural information processing systems , volume=

  10. [10]

    Advances in neural information processing systems , volume=

    Visual instruction tuning , author=. Advances in neural information processing systems , volume=

  11. [11]

    Proceedings of the 25th ACM international conference on Multimedia , pages=

    Video question answering via gradually refined attention over appearance and motion , author=. Proceedings of the 25th ACM international conference on Multimedia , pages=

  12. [12]

    Proceedings of the AAAI conference on artificial intelligence , volume=

    Activitynet-qa: A dataset for understanding complex web videos via question answering , author=. Proceedings of the AAAI conference on artificial intelligence , volume=

  13. [13]

    Proceedings of the 2024 conference on empirical methods in natural language processing , pages=

    Video-llava: Learning united visual representation by alignment before projection , author=. Proceedings of the 2024 conference on empirical methods in natural language processing , pages=

    2024

  14. [14]

    Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , pages=

    Video-chatgpt: Towards detailed video understanding via large vision and language models , author=. Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , pages=

  15. [15]

    International conference on machine learning , pages=

    Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models , author=. International conference on machine learning , pages=. 2023 , organization=

    2023

  16. [16]

    arXiv preprint arXiv:2501.13106 , year=

    Videollama 3: Frontier multimodal foundation models for image and video understanding , author=. arXiv preprint arXiv:2501.13106 , year=

    Pith/arXiv arXiv

  17. [17]

    arXiv preprint arXiv:2511.21631 , year=

    Qwen3-vl technical report , author=. arXiv preprint arXiv:2511.21631 , year=

    Pith/arXiv arXiv

  18. [18]

    arXiv preprint arXiv:2411.10442 , year=

    Enhancing the reasoning ability of multimodal large language models via mixed preference optimization , author=. arXiv preprint arXiv:2411.10442 , year=

    Pith/arXiv arXiv

  19. [19]

    Pattern Recognition , pages=

    SSL-SSAW: Self-Supervised Learning with Sigmoid Self-Attention Weighting for Question-Based Sign Language Translation , author=. Pattern Recognition , pages=. 2026 , publisher=

    2026

  20. [20]

    IEEE Transactions on Circuits and Systems for Video Technology , volume=

    Overcoming modality bias in question-driven sign language video translation , author=. IEEE Transactions on Circuits and Systems for Video Technology , volume=. 2024 , publisher=

    2024

  21. [21]

    arXiv preprint arXiv:2606.03788 , year=

    SLU-2K: A Question-Based Benchmark for Semantic Evaluation of Sign Language Translation , author=. arXiv preprint arXiv:2606.03788 , year=

    Pith/arXiv arXiv

  22. [22]

    2000 , publisher =

    The Syntax of American Sign Language: Functional Categories and Hierarchical Structure , author =. 2000 , publisher =

    2000

  23. [23]

    2006 , publisher =

    Sign Language and Linguistic Universals , author =. 2006 , publisher =

    2006

  24. [24]

    TCSVT , volume=

    Semantic Boundary Detection with Reinforcement Learning for Continuous Sign Language Recognition , author=. TCSVT , volume=. 2020 , publisher=

    2020

  25. [25]

    , author=

    Dilated Convolutional Network with Iterative Optimization for Continuous Sign Language Recognition. , author=. IJCAI , volume=

  26. [26]

    ECCV , year=

    Fully convolutional networks for continuous sign language recognition , author=. ECCV , year=

  27. [27]

    CVPR , pages=

    C2SLR: Consistency-Enhanced Continuous Sign Language Recognition , author=. CVPR , pages=

  28. [28]

    TPAMI , year=

    Weakly supervised learning with multi-stream CNN-LSTM-HMMs to discover sequential parallelism in sign language videos , author=. TPAMI , year=

  29. [29]

    MM , pages=

    Boosting Continuous Sign Language Recognition via Cross Modality Augmentation , author=. MM , pages=

  30. [30]

    ICCV , pages=

    Self-Mutual Distillation Learning for Continuous Sign Language Recognition , author=. ICCV , pages=

  31. [31]

    ICCV , pages=

    Visual alignment constraint for continuous sign language recognition , author=. ICCV , pages=

  32. [32]

    ICIP , pages=

    Continuous sign language recognition via reinforcement learning , author=. ICIP , pages=. 2019 , organization=

    2019

  33. [33]

    ICIMCS , pages=

    A threshold-based hmm-dtw approach for continuous sign language recognition , author=. ICIMCS , pages=

  34. [34]

    CVPR , pages=

    Re-sign: Re-aligned end-to-end sequence modelling with deep recurrent CNN-HMMs , author=. CVPR , pages=

  35. [35]

    AAAI , year=

    Video-based sign language recognition without temporal segmentation , author=. AAAI , year=

  36. [36]

    TMM , year=

    Graph-based multimodal sequential embedding for sign language translation , author=. TMM , year=

  37. [37]

    AAAI , volume=

    Hierarchical lstm for sign language translation , author=. AAAI , volume=

  38. [38]

    ICCV , pages=

    Subunets: End-to-end hand shape and continuous sign language recognition , author=. ICCV , pages=. 2017 , organization=

    2017

  39. [39]

    arXiv preprint arXiv:1908.01341 , year=

    SF-Net: Structured Feature Network for Continuous Sign Language Recognition , author=. arXiv preprint arXiv:1908.01341 , year=

    Pith/arXiv arXiv 1908

  40. [40]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

    SignGraph: A Sign Sequence is Worth Graphs of Nodes , author=. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

  41. [41]

    CVPR , pages=

    Recurrent convolutional neural networks for continuous sign language recognition by staged optimization , author=. CVPR , pages=

  42. [42]

    ECCV , pages=

    Stochastic fine-grained labeling of multi-state sign glosses for continuous sign language recognition , author=. ECCV , pages=. 2020 , organization=

    2020

  43. [43]

    , author=

    Spatial-Temporal Multi-Cue Network for Continuous Sign Language Recognition. , author=. AAAI , year=

  44. [44]

    ICME , pages=

    Dynamic pseudo label decoding for continuous sign language recognition , author=. ICME , pages=. 2019 , organization=

    2019

  45. [45]

    TMM , volume=

    A deep neural framework for continuous sign language recognition by iterative training , author=. TMM , volume=. 2019 , publisher=

    2019

  46. [46]

    CVPR , pages=

    Iterative alignment network for continuous sign language recognition , author=. CVPR , pages=

  47. [47]

    CVIU , pages=

    Koller, Oscar and Forster, Jens and Ney, Hermann , title=. CVIU , pages=

  48. [48]

    Pattern Recognition Letters , volume=

    Continuous sign language recognition using level building based on fast hidden Markov model , author=. Pattern Recognition Letters , volume=. 2016 , publisher=

    2016

  49. [49]

    CVPR , pages=

    Deep hand: How to train a cnn on 1 million hand images when your data is continuous and weakly labelled , author=. CVPR , pages=

  50. [50]

    Proceedings of the British Machine Vision Conference 2016 , year=

    Deep sign: Hybrid CNN-HMM for continuous sign language recognition , author=. Proceedings of the British Machine Vision Conference 2016 , year=

    2016

  51. [51]

    CVPR , pages=

    Continuous Sign Language Recognition with Correlation Network , author=. CVPR , pages=

  52. [52]

    ICCV , pages=

    CoSign: Exploring Co-occurrence Signals in Skeleton-based Continuous Sign Language Recognition , author=. ICCV , pages=

  53. [53]

    ICASSP , year=

    Spatio-temporal graph convolutional networks for continuous sign language recognition , author=. ICASSP , year=

  54. [54]

    Proceedings of the AAAI Conference on Artificial Intelligence , volume=

    Self-emphasizing network for continuous sign language recognition , author=. Proceedings of the AAAI Conference on Artificial Intelligence , volume=

  55. [55]

    Proceedings of the IEEE/CVF International Conference on Computer Vision , pages=

    Improving continuous sign language recognition with cross-lingual signs , author=. Proceedings of the IEEE/CVF International Conference on Computer Vision , pages=

  56. [56]

    Journal of Electronic Imaging , volume=

    Multiscale temporal network for continuous sign language recognition , author=. Journal of Electronic Imaging , volume=. 2024 , publisher=

    2024

  57. [57]

    European conference on computer vision , pages=

    Temporal lift pooling for continuous sign language recognition , author=. European conference on computer vision , pages=. 2022 , organization=

    2022

  58. [58]

    Proceedings of the IEEE/CVF International Conference on Computer Vision , pages=

    A Closer Look at Skeleton-based Continuous Sign Language Recognition , author=. Proceedings of the IEEE/CVF International Conference on Computer Vision , pages=

  59. [59]

    , author=

    Connectionist Temporal Modeling of Video and Language: a Joint Model for Translation and Sign Labeling. , author=. IJCAI , pages=

  60. [60]

    , author=

    Dense Temporal Convolution Network for Sign Language Translation. , author=. IJCAI , pages=

  61. [61]

    ACM Transactions on Accessible Computing (TACCESS) , volume=

    Isolated sign language recognition with grassmann covariance matrices , author=. ACM Transactions on Accessible Computing (TACCESS) , volume=. 2016 , publisher=

    2016

  62. [62]

    Proceedings of the AAAI Conference on Artificial Intelligence , volume=

    Hand-Model-Aware Sign Language Recognition , author=. Proceedings of the AAAI Conference on Artificial Intelligence , volume=

  63. [63]

    IEEE Transactions on Circuits and Systems for Video Technology , volume=

    Attention-based 3D-CNNs for large-vocabulary sign language recognition , author=. IEEE Transactions on Circuits and Systems for Video Technology , volume=. 2018 , publisher=

    2018

  64. [64]

    arXiv preprint arXiv:2008.10428 , year=

    Global-local Enhancement Network for NMFs-aware Sign Language Recognition , author=. arXiv preprint arXiv:2008.10428 , year=

    Pith/arXiv arXiv 2008

  65. [65]

    2018 IEEE International Conference on Big Data (Big Data) , pages=

    American sign language recognition using deep learning and computer vision , author=. 2018 IEEE International Conference on Big Data (Big Data) , pages=. 2018 , organization=

    2018

  66. [66]

    2011 IEEE international conference on computer vision workshops (ICCV Workshops) , pages=

    Real-time sign language letter and word recognition from depth data , author=. 2011 IEEE international conference on computer vision workshops (ICCV Workshops) , pages=. 2011 , organization=

    2011

  67. [67]

    SMC , title=

    K. SMC , title=. 1997 , volume=

    1997

  68. [68]

    Proceedings of the AAAI conference on artificial intelligence , volume=

    BEST: BERT pre-training for sign language recognition with coupling tokenization , author=. Proceedings of the AAAI conference on artificial intelligence , volume=

  69. [69]

    ACM Transactions on Intelligent Systems and Technology (TIST) , volume=

    A real-time hand posture recognition system using deep neural networks , author=. ACM Transactions on Intelligent Systems and Technology (TIST) , volume=. 2015 , publisher=

    2015

  70. [70]

    2016 IEEE International Conference on Image Processing (ICIP) , pages=

    Sign language recognition based on adaptive hmms with data augmentation , author=. 2016 IEEE International Conference on Image Processing (ICIP) , pages=. 2016 , organization=

    2016

  71. [71]

    ICIP , pages=

    Sign language recognition with long short-term memory , author=. ICIP , pages=. 2016 , organization=

    2016

  72. [72]

    TPAMI , year=

    Scaling up multimodal pre-training for sign language understanding , author=. TPAMI , year=

  73. [73]

    Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

    Natural language-assisted sign language recognition , author=. Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

  74. [74]

    Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

    Transferring cross-domain knowledge for video sign language recognition , author=. Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

  75. [75]

    arXiv preprint arXiv:2503.08529 , year=

    Signrep: Enhancing self-supervised sign representations , author=. arXiv preprint arXiv:2503.08529 , year=

    Pith/arXiv arXiv

  76. [76]

    2018 , publisher=

    Addressing the rising prevalence of hearing loss , author=. 2018 , publisher=

    2018

  77. [77]

    Coronavirus , author=

  78. [78]

    Conference on Computer Vision and Pattern Recognition (CVPR) , year=

    Duarte, Amanda and Palaskar, Shruti and Ventura, Lucas and Ghadiyaram, Deepti and DeHaan, Kenneth and Metze, Florian and Torres, Jordi and Giro-i-Nieto, Xavier , title =. Conference on Computer Vision and Pattern Recognition (CVPR) , year=

  79. [79]

    CVPR , volume=

    Online detection and classification of dynamic hand gestures with recurrent 3d convolutional neural networks , author=. CVPR , volume=

  80. [80]

    Proceedings of the 49th annual meeting of the association for computational linguistics: Human language technologies , pages=

    Learning word vectors for sentiment analysis , author=. Proceedings of the 49th annual meeting of the association for computational linguistics: Human language technologies , pages=

Showing first 80 references.