Pith. sign in

REVIEW 4 major objections 6 minor 70 references

Fine-Tuning Video Transformers for Word-Level Bangla Sign Language: A Comparative Analysis for Classification Tasks

T0 review · 4 major / 6 minor · reviewed 2026-08-07 · deepseek-v4-flash

Pith's one-line read Fine-tuning Kinetics-400 pretrained video transformers, especially VideoMAE, achieves 95.5% accuracy on word-level Bangla sign recognition, a new state of the art.

desk verdict Useful empirical benchmark of video transformers for Bangla sign language, but the reporting inconsistencies and thin signer-independent evaluation make the headline numbers hard to trust as-is. read the letter →

arxiv 2506.04367 v1 pith:C6RMBWSZ submitted 2025-06-04 cs.CV

classification cs.CV
keywords Banglasignlanguagevideotransformersfine-tuningisolatedrecognitionMAEsigner-independentevaluationword-leveltransferlearning
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 attempts to establish that pretrained video transformers, when fine-tuned on small and medium Bangla Sign Language (BdSL) vocabularies, set a new performance level for word-level recognition. It reports 95.5% top-1 accuracy on the 60-word BdSLW60 dataset and 81.04% on the front-facing subset of the 401-word BdSLW401 set, markedly above earlier SVM (67.6%) and attention-based BiLSTM (75.1%) baselines. The work also provides the first benchmark results for BdSLW401 and evaluates generalization on LSA64 and WLASL datasets. If these numbers hold under wider signer coverage, they make video transformers the default option for low-resource sign language recognition.

What carries the argument

The load-bearing mechanism is masked video autoencoding (VideoMAE): a tube-masked encoder that processes only about 10% of video tokens, with a lightweight decoder reconstructing the rest during pretraining, then fine-tuned end-to-end for classification. Sampled clips are standardized to fixed length (16 frames at a sample rate of 8 for VideoMAE), resized to 224x224 pixels, and augmented with random cropping, horizontal flipping, and short-side scaling while the pretrained Kinetics-400 weights supply spatial-temporal priors adapted to sign gestures.

What would settle it

Computing per-signer accuracy by leaving out each of the 18 signers one at a time and comparing the averaged result with the U4/U8 split would settle whether the reported 95.5% reflects true signer independence; if holding out any single signer gives much lower accuracy, the headline number overstates generalization.

Watch

Extended reading notes

Core claim

Fine-tuning a Kinetics-400 pretrained VideoMAE model on BdSLW60 yields 95.5% top-1 accuracy, and fine-tuning the same model on the larger BdSLW401 yields 81.04% on the front-facing sign subset. The paper treats this as evidence that masked-video-autoencoder pretraining captures spatiotemporal structure that transfers well to sign language, outperforming the SVM and attention-BiLSTM results that previously defined the state of the art on BdSLW60. The gains are attributed to the model's ability to preserve spatial-temporal features while avoiding the redundancy that hurts other video transformers on short, irregular clips.

Load-bearing premise

The reported signer-independent accuracies assume that the two held-out users used for testing (U4 and U8 on BdSLW60; S04 and S08 on BdSLW401) are representative of unseen signers, since the paper gives no per-user breakdown and uses only those users for the test split.

Editorial extensions

If this is right

  • Word-level BdSL recognition can now be approached with off-the-shelf video transformers rather than handcrafted features or recurrent architectures.
  • BdSLW401, benchmarked here for the first time, gives future work a 401-class target for scaling studies.
  • Frame-rate correction alone hurts accuracy; adding variation to duplicated frames restores the gain, showing that preprocessing details materially change results.
  • Transformer performance degrades on low-quality, imbalanced datasets such as WLASL2000, indicating clear limits when frame quality and class balance are poor.

Reading between the lines

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

  • A natural next test is to leave out each of the 18 signers one at a time and average the accuracies; the paper's two-user split (U4 and U8) leaves that unmeasured, and per-signer numbers would reveal how much the headline accuracy depends on those specific users.
  • The Kinetics-400 pretraining advantage suggests action-recognition priors transfer especially well to sign language, a hypothesis that could be probed by comparing against models pretrained on sign-specific data.
  • The frame-rate-correction-plus-augmentation trick could be applied to other low-frame-rate sign datasets to test whether the accuracy gain is generic or specific to BdSLW60.
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 / 6 minor

Summary. The paper fine-tunes three video transformer architectures (VideoMAE, ViViT, and TimeSformer) on word-level Bangla sign language datasets BdSLW60 and BdSLW401, with additional generalization experiments on LSA64 and WLASL. The authors report signer-independent test accuracies of 95.5% on BdSLW60 and 81.04% on the front-facing BdSLW401 subset for a Kinetics-400-pretrained VideoMAE model, and compare against prior SVM/BiLSTM baselines and public benchmarks. The paper also analyzes the influence of frame rate correction, sampling rate, augmentation, and dataset scale on recognition performance.

Significance. If the reported numbers are reproducible, the work provides a useful baseline for word-level BdSL recognition, which is genuinely under-resourced compared with ASL, Turkish, and Argentinian sign language benchmarks. Strengths include evaluation on three external benchmarks (LSA64, WLASL100, WLASL2000), public release of fine-tuned checkpoints on Hugging Face, and an explicit analysis of frame-rate correction and augmentation. The central claim is plausible because it builds on established video transformers with Kinetics-400 pretraining. However, the significance is currently undercut by internal numerical inconsistencies and by a signer-independent evaluation setup that is too weakly documented to support the headline generalization claims.

major comments (4)
  1. [§4.1, Table 4 and §4.2, Table 5] The same model, 'MCG-NJU/videomae-base-finetuned-kinetics', with augmentation is reported as 92.55% in Table 4 but 95.5% in Table 5 for BdSLW60; the non-finetuned 'videomae-base' model also differs (84.95% vs 93.6%). Since the 95.5% and 81.04% values are the paper's headline claims, the authors must reconcile these numbers, state which configuration each table corresponds to (e.g., 10-fold cross-validation vs the fixed U4/U8 split), and report run-to-run variance or confidence intervals.
  2. [§3.6, Table 3] Signer-independent evaluation rests on only two held-out users (U4/U8 for BdSLW60; S04/S08 for BdSLW401) with no per-user accuracy breakdown. The manuscript also does not state whether the validation user U5 and the 10-fold stratified validation folds are signer-disjoint from the test users; if any validation samples share signer identity with test users, early stopping and checkpoint selection would leak signer identity. Please specify the exact user-to-split mapping for every fold and report per-user test accuracy.
  3. [§1, §3, §4.2, §5] The paper repeatedly claims to provide the first benchmark results on BdSLW401 and states that no prior research has applied video transformers to word-level BdSL recognition, while citing [2] (BdSLW401: Transformer-based word-level Bangla sign language recognition) and [32] (attention-based transformer for BdSL word-level recognition). This is internally contradictory. Please remove or qualify the priority claims and directly compare with the results of [2] and [32] rather than only with SVM and BiLSTM baselines.
  4. [§5, Conclusion] The conclusion states that on BdSLW401 the model achieved 'an F1 score of 80.14%, recall of 84.57%, and precision of 81.14%', which swaps precision and recall relative to Table 5 (Pre 84.57%, Rec 81.14%, F1 80.14%). Along with the Table 4/Table 5 discrepancy, this makes the headline metric reporting unreliable and should be corrected and cross-checked across the abstract, main text, and tables.
minor comments (6)
  1. [Table 4] The row for 'MCG-NJU/videomae-base-finetuned-kinetics' reads '2092.55%' instead of two separate entries '20' and '92.55%'; please fix the formatting.
  2. [Table 5 and §4.2] No TimeSformer results are reported for WLASL2000, although other models are listed; please either add the row or state why it is omitted.
  3. [§4.1] The sentence 'Validation results using VideoMAE and ViViT (Table 4)' is inconsistent with the table columns, which report 'Test Accuracy'; please align the wording.
  4. [§3.4.3, Eq. (1)] The notation z^{(0)}_{(p,t)} and the embedding term 'Ex_{(p,t)}' are not defined; please define the patch and temporal indices and the embedding function.
  5. [References [13]] The LSA64 citation points to a 2023 arXiv abstract (arXiv:2310.17429), but LSA64 is an earlier dataset; the original reference should be cited.
  6. [Introduction] The paper refers to 'Section II', 'Section III', etc., but the manuscript uses numbered sections 1-5; please update the cross-references.

Circularity Check

0 steps flagged · score 1.0 of 10

No circular derivation; empirical fine-tuning results are self-contained, with only minor non-load-bearing self-citation.

full rationale

The paper contains no mathematical derivation chain in which a predicted quantity is defined in terms of the target or a fitted parameter is renamed as a prediction. The headline accuracies (95.5% on BdSLW60 and 81.04% on BdSLW401-front) are obtained by fine-tuning externally pretrained checkpoints (MCG-NJU/videomae-base-finetuned-kinetics, google/vivit, facebook/TimeSformer) on held-out user splits (U4/U8 for BdSLW60, S04/S08 for BdSLW401) and by benchmarking on external datasets (LSA64, WLASL), so they are not forced by construction. The only self-citation-adjacent elements are the use of the authors' own prior BdSLW60 baselines [1] and the assertion of a 'first benchmark' on BdSLW401 while citing the authors' own prior work [2]; neither is load-bearing for the accuracy values, which are additionally compared against externally published LSA64 and WLASL results. Reporting weaknesses—no per-user accuracy breakdown, possible ambiguity about fold/test-user disjointness in the 10-fold setup, and swapped precision/recall in the conclusion—are correctness and reproducibility concerns, not circularity. Accordingly the circularity burden is minimal.

Assumptions & free parameters 2 free parameters · 4 assumptions · 0 invented entities

The empirical claims rely on a small set of hand-chosen sampling and preprocessing values plus dataset-quality and transferability assumptions. No new physical or architectural entities are introduced.

free parameters (2)
  • Per-dataset sampling rate SR = BdSLW60: VideoMAE 8, ViViT 4, TimeSformer 16; BdSLW401: 10, 5, 20; LSA64: 12, 6, 24; WLASL: 8, 4, 16
    Chosen by hand per model architecture and dataset (Table 6); determines which frames are seen and directly affects accuracy, so it functions as a tuned input parameter rather than a derived quantity.
  • Clip duration and padding length = 4.27 s (BdSLW60), 5.34 s (BdSLW401), 3.2 s (LSA64), 5.2 s (WLASL)
    Computed as C = N*S/F once the frame count and sampling rate are fixed; the choice of padding shorter clips to the longest duration is an ad hoc preprocessing decision that can influence temporal modeling.
assumptions (4)
  • domain assumption Pretrained Kinetics-400 video representations transfer to sign language gestures
    The fine-tuning recipe assumes the source task (human action recognition) provides useful spatiotemporal features for BdSL; the paper motivates this in Section 2 but does not demonstrate it separately.
  • domain assumption Upsampling 15/24 FPS videos to 30 FPS by frame duplication preserves the linguistic content needed for classification
    Section 4.3 admits FRC alone degraded accuracy, so the correction plus augmentation is assumed to restore usable temporal information.
  • domain assumption The class labels and clip boundaries in BdSLW60 and BdSLW401 are correct
    Dataset quality is taken from the source papers [1,2]; no manual verification is reported.
  • domain assumption Held-out users U4, U8 (and S04, S08 for BdSLW401) are representative of unseen signers
    The signer-independent claim rests on these two users; no per-user breakdown or demographic information is provided.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Fine-Tuning Video Transformers for Word-Level Bangla Sign Language: A Comparative Analysis for Classification Tasks." pith.science (2026). https://pith.science/paper/C6RMBWSZ

@misc{pith2026250604367,
  author       = {Pith},
  title        = {Pith review of: Fine-Tuning Video Transformers for Word-Level Bangla Sign Language: A Comparative Analysis for Classification Tasks},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/C6RMBWSZ}},
  note         = {Machine review of arXiv:2506.04367}
}
read the original abstract

Sign Language Recognition (SLR) involves the automatic identification and classification of sign gestures from images or video, converting them into text or speech to improve accessibility for the hearing-impaired community. In Bangladesh, Bangla Sign Language (BdSL) serves as the primary mode of communication for many individuals with hearing impairments. This study fine-tunes state-of-the-art video transformer architectures -- VideoMAE, ViViT, and TimeSformer -- on BdSLW60 (arXiv:2402.08635), a small-scale BdSL dataset with 60 frequent signs. We standardized the videos to 30 FPS, resulting in 9,307 user trial clips. To evaluate scalability and robustness, the models were also fine-tuned on BdSLW401 (arXiv:2503.02360), a large-scale dataset with 401 sign classes. Additionally, we benchmark performance against public datasets, including LSA64 and WLASL. Data augmentation techniques such as random cropping, horizontal flipping, and short-side scaling were applied to improve model robustness. To ensure balanced evaluation across folds during model selection, we employed 10-fold stratified cross-validation on the training set, while signer-independent evaluation was carried out using held-out test data from unseen users U4 and U8. Results show that video transformer models significantly outperform traditional machine learning and deep learning approaches. Performance is influenced by factors such as dataset size, video quality, frame distribution, frame rate, and model architecture. Among the models, the VideoMAE variant (MCG-NJU/videomae-base-finetuned-kinetics) achieved the highest accuracies of 95.5% on the frame rate corrected BdSLW60 dataset and 81.04% on the front-facing signs of BdSLW401 -- demonstrating strong potential for scalable and accurate BdSL recognition.

Figures

Figures reproduced from arXiv: 2506.04367 by the authors.

Figure 1
Figure 1. Architecture of Frame Rate-Corrected Dataset Construction, Recognition, and Benchmarking. [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. Frame Count vs Number of short clips of BdSLW60 dataset [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. VideoMAE Architecture for BdSLW60 Dataset Recognition [53] [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figures from the paper (5 more)
Figure 4
Figure 4. Figure 4: Structure of the ViViT Model and Its Variations for BdSLW60 Recognition [54] [PITH_FULL_IMAGE:figures/full_fig_p006_4.png]
Figure 5
Figure 5. Figure 5: Five Self-Attention Blocks of TimeSformer [55] [PITH_FULL_IMAGE:figures/full_fig_p007_5.png]
Figure 6
Figure 6. Figure 6: Loss Curve for Fold 9 of the BdSLW60 Dataset [PITH_FULL_IMAGE:figures/full_fig_p011_6.png]
Figure 7
Figure 7. Figure 7: Confusion Matrix for Fold 9 of the BdSLW60 Test Set [PITH_FULL_IMAGE:figures/full_fig_p011_7.png]
Figure 8
Figure 8. Figure 8: Confusion Matrix for BdSLW401 Test Set (Visualizing First 50 Classes) [PITH_FULL_IMAGE:figures/full_fig_p012_8.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

70 extracted references · 56 canonical work pages

  1. [2]

    Bdslw401: Transformer- based word-level bangla sign language recognition using relative quantization encoding (rqe).arXiv preprint arXiv:2503.02360, 2025

    Husne Ara Rubaiyeat, Njayou Youssouf, Md Kamrul Hasan, and Hasan Mahmud. Bdslw401: Transformer- based word-level bangla sign language recognition using relative quantization encoding (rqe).arXiv preprint arXiv:2503.02360, 2025

  2. [32]

    Kazi Reyazul Hasan and Muhammad Abdullah Adnan. Empath: Mediapipe-aided ensemble learning with attention- based transformers for accurate recognition of bangla word-level sign language.International Conference on Pattern Recognition, pages 355–371, 2025

  3. [1]

    Bdslw60: A word-level bangla sign language dataset.Multimedia Tools and Applications, pages 1–25, 2025

    Husne Ara Rubaiyeat, Hasan Mahmud, Ahsan Habib, and Md Kamrul Hasan. Bdslw60: A word-level bangla sign language dataset.Multimedia Tools and Applications, pages 1–25, 2025

  4. [3]

    Deafness and hearing loss

    World Health Organization. Deafness and hearing loss. 2021. Available: https://www.who.int/ news-room/fact-sheets/detail/deafness-and-hearing-loss

  5. [4]

    Sneha Sharma, Rinki Gupta, and A. Kumar. Continuous sign language recognition using isolated signs data and deep transfer learning.Journal of Ambient Intelligence and Humanized Computing, 14:1–12, 08 2021

  6. [5]

    Biyi Fang, Jillian Co, and Mi Zhang. Deepasl: Enabling ubiquitous and non-intrusive word and sentence-level sign language translation.Proceedings of the 15th ACM conference on embedded network sensor systems, pages 1–13, 2017. 13 APREPRINT- SEPTEMBER13, 2025

  7. [6]

    Recent advances on deep learning for sign language recognition.CMES - Computer Modeling in Engineering and Sciences, 139(3):2399–2450, 2024

    Yanqiong Zhang and Xianwei Jiang. Recent advances on deep learning for sign language recognition.CMES - Computer Modeling in Engineering and Sciences, 139(3):2399–2450, 2024

  8. [7]

    Deeparslr: A novel signer-independent deep learning framework for isolated arabic sign language gestures recognition.IEEE Access, 8:83199–83212, 2020

    Saleh Aly and Walaa Aly. Deeparslr: A novel signer-independent deep learning framework for isolated arabic sign language gestures recognition.IEEE Access, 8:83199–83212, 2020

Show all 70 references
  1. [8]

    Kishore, Teja Maddala, and D

    Kiran Eepuri, P.V .V . Kishore, Teja Maddala, and D. Anil Kumar. 3d sign language recognition with joint distance and angular coded color topographical descriptor on a 2 – stream cnn.Neurocomputing, 372, 09 2019

  2. [9]

    Using motion history images with 3d convolutional networks in isolated sign language recognition.IEEE Access, 10:18608–18618, 2022

    Ozge Mercanoglu Sincan and Hacer Yalim Keles. Using motion history images with 3d convolutional networks in isolated sign language recognition.IEEE Access, 10:18608–18618, 2022

  3. [10]

    Alba-Castro, Laura Docio-Fernandez, and Eduardo Rodriguez-Banga

    Manuel Vazquez-Enriquez, Jose L. Alba-Castro, Laura Docio-Fernandez, and Eduardo Rodriguez-Banga. Isolated sign language recognition with multi-scale spatial-temporal graph convolutional networks.Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition...

  4. [11]

    One model is not enough: Ensembles for isolated sign language recognition.Sensors, 22(13), 2022

    Marek Hrúz, Ivan Gruber, Jakub Kanis, Matyáš Boháˇcek, Miroslav Hlaváˇc, and Zdenˇek Krˇnoul. One model is not enough: Ensembles for isolated sign language recognition.Sensors, 22(13), 2022

  5. [12]

    David Laines, Miguel Gonzalez-Mendoza, Gilberto Ochoa-Ruiz, and Gissella Bejarano. Isolated sign language recognition based on tree structure skeleton images.Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, pages 276–284, June 2023

  6. [13]

    Lsa64: An argentinian sign language dataset.arXiv preprint arXiv:2310.17429, 2023

    Franco Ronchetti, Facundo Manuel Quiroga, César Estrebou, Laura Lanzarini, and Alejandro Rosete. Lsa64: An argentinian sign language dataset.arXiv preprint arXiv:2310.17429, 2023

  7. [14]

    Best: Bert pre-training for sign language recognition with coupling tokenization.Proceedings of the AAAI conference on artificial intelligence, 37(3):3597–3605, 2023

    Weichao Zhao, Hezhen Hu, Wengang Zhou, Jiaxin Shi, and Houqiang Li. Best: Bert pre-training for sign language recognition with coupling tokenization.Proceedings of the AAAI conference on artificial intelligence, 37(3):3597–3605, 2023

  8. [15]

    Deep learning for computer vision: A brief review.Intell

    Athanasios V oulodimos, Nikolaos Doulamis, Anastasios Doulamis, Eftychios Protopapadakis, and Diego Andina. Deep learning for computer vision: A brief review.Intell. Neuroscience, 2018, January 2018

  9. [16]

    Asl-3dcnn: American sign language recognition technique using 3-d convolutional neural networks.Multimedia Tools and Applications, 80(17):26319–26331, 2021

    Shikhar Sharma and Krishan Kumar. Asl-3dcnn: American sign language recognition technique using 3-d convolutional neural networks.Multimedia Tools and Applications, 80(17):26319–26331, 2021

  10. [17]

    3d convolutional neural networks for dynamic sign language recognition.The Computer Journal, 61(11):1724–1736, 2018

    Zhi-jie Liang, Sheng-bin Liao, and Bing-zhang Hu. 3d convolutional neural networks for dynamic sign language recognition.The Computer Journal, 61(11):1724–1736, 2018

  11. [18]

    Fine-tuning of sign language recognition models: a technical report.arXiv preprint arXiv:2302.07693, 2023

    Maxim Novopoltsev, Leonid Verkhovtsev, Ruslan Murtazin, Dmitriy Milevich, and Iuliia Zemtsova. Fine-tuning of sign language recognition models: a technical report.arXiv preprint arXiv:2302.07693, 2023

  12. [19]

    Self- supervised video transformers for isolated sign language recognition.arXiv preprint arXiv:2309.02450, 2023

    Marcelo Sandoval-Castaneda, Yanhong Li, Diane Brentari, Karen Livescu, and Gregory Shakhnarovich. Self- supervised video transformers for isolated sign language recognition.arXiv preprint arXiv:2309.02450, 2023

  13. [20]

    Full transformer network with masking future for word-level sign language recognition.Neurocomputing, 500:115–123, 2022

    Yao Du, Pan Xie, Mingye Wang, Xiaohui Hu, Zheng Zhao, and Jiaqi Liu. Full transformer network with masking future for word-level sign language recognition.Neurocomputing, 500:115–123, 2022

  14. [21]

    Sign language recognition from digital videos using feature pyramid network with detection transformer.Multimedia Tools and Applications, 82(14):21673–21685, 2023

    Yu Liu, Parma Nand, Md Akbar Hossain, Minh Nguyen, and Wei Qi Yan. Sign language recognition from digital videos using feature pyramid network with detection transformer.Multimedia Tools and Applications, 82(14):21673–21685, 2023

  15. [22]

    Yutong Chen, Fangyun Wei, Xiao Sun, Zhirong Wu, and Stephen Lin. A simple multi-modality transfer learning baseline for sign language translation.Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5120–5130, June 2022

  16. [23]

    Transfer learning in sign language.2007 IEEE conference on computer vision and pattern recognition, pages 1–8, 2007

    Ali Farhadi, David Forsyth, and Ryan White. Transfer learning in sign language.2007 IEEE conference on computer vision and pattern recognition, pages 1–8, 2007

  17. [24]

    Samir Imtiaz, Nieb Hasan Neom, Nazmul Siddique, and Hui Wang

    Sunanda Das, Md. Samir Imtiaz, Nieb Hasan Neom, Nazmul Siddique, and Hui Wang. A hybrid approach for bangla sign language recognition using deep transfer learning model with random forest classifier.Expert Systems with Applications, 213:118914, 2023

  18. [25]

    Bird, Anikó Ekárt, and Diego R

    Jordan J. Bird, Anikó Ekárt, and Diego R. Faria. British sign language recognition via late fusion of computer vision and leap motion with transfer learning to american sign language.Sensors, 20(18), 2020

  19. [26]

    Transfer learning for british sign language modelling.arXiv preprint arXiv:2006.02144, 2020

    Boris Mocialov, Graham Turner, and Helen Hastie. Transfer learning for british sign language modelling.arXiv preprint arXiv:2006.02144, 2020

  20. [27]

    Alharthi and Salha M

    Nojood M. Alharthi and Salha M. Alzahrani. Vision transformers and transfer learning approaches for arabic sign language recognition.Applied Sciences, 13(21), 2023. 14 APREPRINT- SEPTEMBER13, 2025

  21. [28]

    Multi-phase fine-tuning: A new fine-tuning approach for sign language recognition.KI - Künstliche Intelligenz, 36, 02 2022

    Noha Sarhan, Mikko Lauri, and Simone Frintrop. Multi-phase fine-tuning: A new fine-tuning approach for sign language recognition.KI - Künstliche Intelligenz, 36, 02 2022

  22. [29]

    Khalil Ahammad, Jubayer Ahmed Bhuiyan Shawon, Partha Chakraborty, Md Jahidul Islam, and Saiful Islam. Recognizing bengali sign language gestures for digits in real time using convolutional neural network.International Journal of Computer Science and Information Security (IJCSI...

  23. [30]

    Kanchon Kanti Podder, Muhammad E. H. Chowdhury, Anas M. Tahir, Zaid Bin Mahbub, Amith Khandakar, Md Shafayet Hossain, and Muhammad Abdul Kadir. Bangla sign language (bdsl) alphabets and numerals classification using a deep learning model.Sensors, 22(2), 2022

  24. [31]

    Borno-net: A real-time bengali sign-character detection and sentence generation system using quantized yolov4-tiny and lstms.Applied Sciences, 13(9), 2023

    Nasima Begum, Rashik Rahman, Nusrat Jahan, Saqib Sizan Khan, Tanjina Helaly, Ashraful Haque, and Nipa Khatun. Borno-net: A real-time bengali sign-character detection and sentence generation system using quantized yolov4-tiny and lstms.Applied Sciences, 13(9), 2023

  25. [33]

    Computer vision-based hybrid efficient convolution for isolated dynamic sign language recognition.Neural Computing and Applications, 36(32):19951–19966, 2024

    Prothoma Khan Chowdhury, Kabiratun Ummi Oyshe, Muhammad Aminur Rahaman, Tanoy Debnath, Anichur Rahman, and Neeraj Kumar. Computer vision-based hybrid efficient convolution for isolated dynamic sign language recognition.Neural Computing and Applications, 36(32):19951–19966, 2024

  26. [34]

    Madhiarasan and Partha Pratim Roy

    M. Madhiarasan and Partha Pratim Roy. A comprehensive review of sign language recognition: Different types, modalities, and datasets.arXiv preprint arXiv:2204.03328, 2022

  27. [35]

    A review of hand gesture and sign language recognition techniques.International Journal of Machine Learning and Cybernetics, 10, 01 2019

    Ming Jin Cheok, Zaid Omar, and Mohamed Jaward. A review of hand gesture and sign language recognition techniques.International Journal of Machine Learning and Cybernetics, 10, 01 2019

  28. [36]

    Recognition of symbolic gestures using depth information.Advances in Human-Computer Interaction, 2018(1):1069823, 2018

    Hasan Mahmud, Md Kamrul Hasan, Md Hasanul Kabir, and MA Mottalib. Recognition of symbolic gestures using depth information.Advances in Human-Computer Interaction, 2018(1):1069823, 2018

  29. [37]

    AL-Rousan, K

    M. AL-Rousan, K. Assaleh, and A. Tala’a. Video-based signer-independent arabic sign language recognition using hidden markov models.Applied Soft Computing, 9(3):990–999, 2009

  30. [38]

    Signer independent isolated italian sign recognition based on hidden markov models.Formal Pattern Analysis & Applications, 18, 10 2014

    Marco Fagiani, Emanuele Principi, Stefano Squartini, and Francesco Piazza. Signer independent isolated italian sign recognition based on hidden markov models.Formal Pattern Analysis & Applications, 18, 10 2014

  31. [39]

    Rabeet Fatmi, Sherif Rashad, and Ryan Integlia. Comparing ann, svm, and hmm based machine learning methods for american sign language recognition using wearable motion sensors.2019 IEEE 9th Annual Computing and Communication Workshop and Conference (CCWC), pages 0290–0297, 2019

  32. [40]

    Sarfaraz Masood, Adhyan Srivastava, Harish Chandra Thuwal, and Musheer Ahmad. Real-time sign language gesture (word) recognition from video sequences using cnn and rnn.Intelligent Engineering Informatics: Proceedings of the 6th International Conference on FICTA, pages 623–632, 2018

  33. [41]

    Video-based isolated hand sign language recognition using a deep cascaded model.Multimedia Tools and Applications, 79, 08 2020

    Razieh Rastgoo, Kourosh Kiani, and Sergio Escalera. Video-based isolated hand sign language recognition using a deep cascaded model.Multimedia Tools and Applications, 79, 08 2020

  34. [42]

    Hand pose aware multimodal isolated sign language recognition.Multimedia Tools and Applications, 80(1):127–163, 2021

    Razieh Rastgoo, Kourosh Kiani, and Sergio Escalera. Hand pose aware multimodal isolated sign language recognition.Multimedia Tools and Applications, 80(1):127–163, 2021

  35. [43]

    Al Mehedi Hasan, Koki Hirooka, Kota Suzuki, Hyoun-Sup Lee, and Si-Woong Jang

    Jungpil Shin, Abu Saleh Musa Miah, Md. Al Mehedi Hasan, Koki Hirooka, Kota Suzuki, Hyoun-Sup Lee, and Si-Woong Jang. Korean sign language recognition using transformer-based deep neural network.Applied Sciences, 13(5), 2023

  36. [44]

    Johir Raihan, Mainul Islam Labib, Abdullah Al Jaid Jim, Jun Jiat Tiang, Uzzal Biswas, and Abdullah-Al Nahid

    Md. Johir Raihan, Mainul Islam Labib, Abdullah Al Jaid Jim, Jun Jiat Tiang, Uzzal Biswas, and Abdullah-Al Nahid. Bengali-sign: A machine learning-based bengali sign language interpretation for deaf and non-verbal people.Sensors, 24(16), 2024

  37. [45]

    Sign language recognition with transformer networks.Proceedings of the Twelfth Language Resources and Evaluation Conference, pages 6018–6024, 2020

    Mathieu De Coster, Mieke Van Herreweghe, and Joni Dambre. Sign language recognition with transformer networks.Proceedings of the Twelfth Language Resources and Evaluation Conference, pages 6018–6024, 2020

  38. [46]

    Sarah Alyami, Hamzah Luqman, and Mohammad Hammoudeh. Isolated arabic sign language recognition using a transformer-based model and landmark keypoints.ACM Transactions on Asian and Low-Resource Language Information Processing, 23(1):1–19, 2024

  39. [47]

    Bdsl47: A complete depth-based bangla sign alphabet and digit dataset.Data in Brief, 51:109799, 2023

    S M Rayeed, Sidratul Tamzida Tuba, Hasan Mahmud, Mumtahin Habib Ullah Mazumder, Saddam Hossain Mukta, and Kamrul Hasan. Bdsl47: A complete depth-based bangla sign alphabet and digit dataset.Data in Brief, 51:109799, 2023

  40. [48]

    Efficient 3d cnns with knowledge transfer for sign language recognition

    Xiangzu Han, Fei Lu, and Guohui Tian. Efficient 3d cnns with knowledge transfer for sign language recognition. Multimedia Tools and Applications, 81(7):10071–10090, 2022. 15 APREPRINT- SEPTEMBER13, 2025

  41. [49]

    Bangla sign alphabet recognition with zero-shot and transfer learning.Pattern Recognition Letters, 150:84–93, 2021

    Ragib Amin Nihal, Sejuti Rahman, Nawara Mahmood Broti, and Shamim Ahmed Deowan. Bangla sign alphabet recognition with zero-shot and transfer learning.Pattern Recognition Letters, 150:84–93, 2021

  42. [50]

    Abdullah Al Rafi, Rakibul Hassan, Md Rabiul Islam, and Md Nahiduzzaman. Real-time lightweight bangla sign language recognition model using pre-trained mobilenetv2 and conditional dcgan.Proceedings of International Conference on Information and Communication Technology for Deve...

  43. [51]

    A short video classification framework based on cross-modal fusion.Sensors, 23(20), 2023

    Nuo Pang, Songlin Guo, Ming Yan, and Chien Aun Chan. A short video classification framework based on cross-modal fusion.Sensors, 23(20), 2023

  44. [52]

    Dodge, and Mohammad H

    Jian Sun, Hiroko H. Dodge, and Mohammad H. Mahoor. Mc-vivit: Multi-branch classifier-vivit to detect mild cognitive impairment in older adults using facial videos.Expert Systems with Applications, 238:121929, 2024

  45. [53]

    Videomae: Masked autoencoders are data-efficient learners for self-supervised video pre-training.arXiv preprint arXiv:2203.12602, 2022

    Zhan Tong, Yibing Song, Jue Wang, and Limin Wang. Videomae: Masked autoencoders are data-efficient learners for self-supervised video pre-training.arXiv preprint arXiv:2203.12602, 2022

  46. [54]

    Vivit: A video vision transformer.arXiv preprint arXiv:2103.15691, 2021

    Anurag Arnab, Mostafa Dehghani, Georg Heigold, Chen Sun, Mario Luˇci´c, and Cordelia Schmid. Vivit: A video vision transformer.arXiv preprint arXiv:2103.15691, 2021

  47. [55]

    Is space-time attention all you need for video understanding? arXiv preprint arXiv:2102.05095, 2021

    Gedas Bertasius, Heng Wang, and Lorenzo Torresani. Is space-time attention all you need for video understanding? arXiv preprint arXiv:2102.05095, 2021

  48. [56]

    The kinetics human action video dataset.arXiv preprint arXiv:1705.06950, 2017

    Will Kay, Joao Carreira, Karen Simonyan, Brian Zhang, Chloe Hillier, Sudheendra Vijayanarasimhan, Fabio Viola, Tim Green, Trevor Back, Paul Natsev, et al. The kinetics human action video dataset.arXiv preprint arXiv:1705.06950, 2017

  49. [57]

    Masked autoencoders are scalable vision learners.arXiv preprint arXiv:2111.06377, 2021

    Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollár, and Ross Girshick. Masked autoencoders are scalable vision learners.arXiv preprint arXiv:2111.06377, 2021

  50. [58]

    Mgmae: Motion guided masking for video masked autoencoding.Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13493–13504, 2023

    Bingkun Huang, Zhiyu Zhao, Guozhen Zhang, Yu Qiao, and Limin Wang. Mgmae: Motion guided masking for video masked autoencoding.Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13493–13504, 2023

  51. [59]

    An image is worth 16x16 words: Transformers for image recognition at scale.arXiv preprint arXiv:2010.11929, 2021

    Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at...

  52. [60]

    Quo vadis, action recognition? a new model and the kinetics dataset

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

  53. [61]

    X3d: Expanding architectures for efficient video recognition.Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2020

    Christoph Feichtenhofer. X3d: Expanding architectures for efficient video recognition.Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2020

  54. [62]

    Video action transformer network.arXiv preprint arXiv:1812.02707, 2019

    Rohit Girdhar, João Carreira, Carl Doersch, and Andrew Zisserman. Video action transformer network.arXiv preprint arXiv:1812.02707, 2019

  55. [63]

    Non-local neural networks.arXiv preprint arXiv:1711.07971, 2018

    Xiaolong Wang, Ross Girshick, Abhinav Gupta, and Kaiming He. Non-local neural networks.arXiv preprint arXiv:1711.07971, 2018

  56. [64]

    Long-term feature banks for detailed video understanding.arXiv preprint arXiv:1812.05038, 2019

    Chao-Yuan Wu, Christoph Feichtenhofer, Haoqi Fan, Kaiming He, Philipp Krähenbühl, and Ross Girshick. Long-term feature banks for detailed video understanding.arXiv preprint arXiv:1812.05038, 2019

  57. [65]

    Attention is all you need.Advances in neural information processing systems, 30, 2017

    Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Łukasz Kaiser, and Illia Polosukhin. Attention is all you need.Advances in neural information processing systems, 30, 2017

  58. [66]

    Fine-tuning cnn image retrieval with no human annotation

    Filip Radenovi´c, Giorgos Tolias, and Ondˇrej Chum. Fine-tuning cnn image retrieval with no human annotation. IEEE Transactions on Pattern Analysis and Machine Intelligence, 41(7):1655–1668, 2019

  59. [67]

    Arabic sign language letters recognition using vision transformer

    Aya F Alnabih and Ashraf Y Maghari. Arabic sign language letters recognition using vision transformer. Multimedia Tools and Applications, 83(34):81725–81739, 2024

  60. [68]

    Word-level deep sign language recognition from video: A new large-scale dataset and methods comparison.arXiv preprint arXiv:1910.11006, 2020

    Dongxu Li, Cristian Rodriguez Opazo, Xin Yu, and Hongdong Li. Word-level deep sign language recognition from video: A new large-scale dataset and methods comparison.arXiv preprint arXiv:1910.11006, 2020

  61. [69]

    Bencherif, Mansour Alsulaiman, Ghulam Muhammad, Mohamed Amine Mekhtiche, Wadood Abdul, Yousef A

    Muneer Al-Hammadi, Mohamed A. Bencherif, Mansour Alsulaiman, Ghulam Muhammad, Mohamed Amine Mekhtiche, Wadood Abdul, Yousef A. Alohali, Tareq S. Alrayes, Hassan Mathkour, Mohammed Faisal, Mo- hammed Algabri, Hamdi Altaheri, Taha Alfakih, and Hamid Ghaleb. Spatial attention-bas...

  62. [70]

    Hierarchical windowed graph attention network and a large scale dataset for isolated indian sign language recognition

    S Patra, A Maitra, M Tiwari, K Kumaran, S Prabhu, S Punyeshwarananda, and S Samanta. Hierarchical windowed graph attention network and a large scale dataset for isolated indian sign language recognition. arxiv 2024.arXiv preprint arXiv:2407.14224, 2024. 16

Pith tools

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