Pith. sign in

REVIEW 3 major objections 8 minor 39 references

Frame-Level Real-Time Assessment of Stroke Rehabilitation Exercises from Video-Level Labeled Data: Task-Specific vs. Foundation Models

T0 review · 3 major / 8 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read Frame-level stroke-rehab classifiers can be trained from video-level labels alone; Action Transformer + Integrated Gradients reaches AUC 72%, above the 69% from ground-truth frame labels.

desk verdict A useful comparative study with an overstated headline claim: the 72% vs 69% frame-level difference is not statistically significant per their own Table IV, and the threshold selection protocol is undocumented. read the letter →

arxiv 2506.03752 v1 pith:4UKTYOHH submitted 2025-06-04 eess.IV

Gonçalo Mesquita , Ana Rita Cóias , Artur Dubrawski , Alexandre Bernardino This is my paper
classification eess.IV
keywords strokerehabilitationcompensatorymotionframe-levelassessmentweaklysupervisedlearningsaliencymapsintegratedgradientspseudo-labelingpose-basedactionrecognition
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

This paper tries to show that real-time, frame-level detection of compensatory motions in stroke rehabilitation exercises can be trained using only video-level labels, without expensive frame-by-frame annotation by therapists. The authors build pseudo-labels for individual frames by taking a video-level classifier, computing input-gradient saliency maps, and thresholding per-frame saliency scores; those pseudo-labels then train a frame-level MLP. On the SERE dataset of 18 post-stroke patients, the best configuration—the Action Transformer with Integrated Gradients—reaches a frame-level AUC of 72%, numerically above the 69% of an MLP trained on ground-truth frame labels. If this holds, a virtual coach could give per-frame feedback with less annotation cost and adapt to new patients more cheaply.

What carries the argument

The load-bearing mechanism is a two-stage pseudo-label pipeline. A video-level classifier (Model A) is trained on body-pose keypoints with video-level trial labels; after a trial is predicted as compensatory, the gradient of the prediction with respect to the input features is computed—either Vanilla Gradient or Integrated Gradients—and the per-frame gradient magnitudes are aggregated, min-max normalized to pseudo-scores $s_i^t$, and thresholded by a single threshold or two uncertain-band thresholds to assign each frame a pseudo-label $z_i^t$. These pseudo-labels, not ground-truth frame labels, train a frame-level MLP (Model B) whose predictions are evaluated against manual frame annotations under leave-one-subject-out cross-validation. The threshold scheme is what converts video-level supervision into the frame-level supervision that an interactive coach would need.

What would settle it

Randomly permute the per-frame saliency scores within each compensatory trial before thresholding and retrain the frame-level MLP; if the frame-level AUC stays near 0.72, the pseudo-labels carry no temporal localization and the result is an artifact of video-level label distribution rather than frame-level compensation detection.

Watch

Extended reading notes

Core claim

The central claim is that frame-level assessment of compensatory movements can be learned from video-level labels alone, and that pretrained models make this work better than a task-specific LSTM trained from scratch. The authors fine-tune four video-level classifiers—LSTM, Action Transformer, SkateFormer, and MOMENT—on pose keypoints with video-level labels, use Vanilla or Integrated Gradients to produce per-frame saliency maps, normalize and threshold those scores into frame pseudo-labels, and train an MLP frame classifier on the pseudo-labels. MOMENT gives the best video-level result (AUC 73% vs 58% for the LSTM), but the Action Transformer with Integrated Gradients gives the best frame-level result (AUC 72%), slightly beating the MLP trained with ground-truth frame labels (AUC 69%). The authors interpret this as evidence that saliency-derived pseudo-labels can be more reliable than human frame annotations, which are subjective and noisy, and that pretrained models improve generalization to held-out patients.

Load-bearing premise

The load-bearing premise is that high saliency scores on frames of a compensatory trial actually mark the moments when the compensatory motion occurs, and low scores mark normal movement; if the gradients instead highlight joints or entire trials, the pseudo-labels are noise and the frame-level AUC no longer measures compensation detection.

Editorial extensions

If this is right

  • Frame-level annotations are no longer required to train a real-time frame-level compensational-motion detector; only video-level trial labels are needed.
  • Pretrained action-recognition and time-series models transfer better to new stroke patients than a from-scratch LSTM, with MOMENT reaching AUC 73% at the video level.
  • Integrated Gradients pseudo-labels train a frame-level classifier that matches or exceeds one trained on ground-truth frame labels, indicating that manual frame annotations are noisy enough that automatically generated labels can compete with them.
  • The choice between single and dual thresholding did not change the Action Transformer's frame-level performance, while dual thresholding helped SkateFormer.

Reading between the lines

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

  • If saliency pseudo-labels are genuinely locating compensation in time, the same recipe—video-level labels plus integrated-gradient thresholding—could transfer to other clinical movement assessments where only session-level labels exist, such as gait freezing or ataxia rating; a direct test would be whether the derived frame boundaries match clinician-marked events.
  • The 72% vs 69% result may partly reflect noise in the human frame labels used as ground truth; if so, evaluating the method against a more reliable annotation protocol, such as majority voting by several therapists, could change the comparison.
  • The method presupposes that compensatory motion occupies a bounded part of a trial; for exercises where compensation is sustained across most of the movement, the saliency threshold would label almost everything positive and the frame-level AUC would overstate localization accuracy.
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 / 8 minor

Summary. The paper proposes a weakly supervised framework for frame-level assessment of compensatory motions in stroke rehabilitation exercises, using only video-level labels. Body pose keypoints are extracted with MediaPipe, a video-level classifier (LSTM baseline, AcT, SkateFormer, or the MOMENT foundation model) is fine-tuned under leave-one-subject-out (LOSO) cross-validation, and frame-level pseudo-labels are generated by thresholding min-max normalized saliency scores from Vanilla Gradient or Integrated Gradients. These pseudo-labels train a frame-level MLP classifier, evaluated against manual frame-level ground truth on the SERE dataset (18 patients, five exercises). The paper reports that MOMENT achieves the best video-level AUC (73% vs. 58% for LSTM) and that AcT with Integrated Gradients achieves the best frame-level AUC (72% vs. 69% for a ground-truth-trained baseline), concluding that the approach can reduce annotation cost while matching or exceeding manual labeling quality.

Significance. If validated, the framework addresses a practically important problem: training real-time frame-level classifiers without expensive manual frame annotations for stroke rehabilitation. The study has notable strengths: evaluation is performed at the patient level with LOSO cross-validation, multiple pretrained model families are compared, two saliency attribution techniques and two pseudo-label selection strategies are explored, and the frame-level evaluation uses AUC, which is threshold-independent. However, the paper's central statistical claims are not supported by the reported tests: the key frame-level comparison (72% vs. 69%) is marked as not statistically significant, and the video-level superiority of MOMENT over the task-specific models is also not significant. In addition, the pseudo-label threshold selection protocol is under-specified, creating a potential data-leak risk. The underlying idea is promising and the experiments are relevant, but the current reporting does not substantiate the headline conclusions.

major comments (3)
  1. [Section III-B, Table II] The pseudo-label thresholds in Table II are described as being 'determined for each model to balance the ratio between False Positives and False Negatives,' but the paper does not state which data were used for this tuning. If the thresholds were selected on the full SERE dataset—including the held-out patient in each LOSO fold—then the frame-level pseudo-labels used to train Model B depend on test-patient frames, and the comparison in Table IV is no longer a valid weakly supervised evaluation. Please specify the exact protocol: were thresholds fixed a priori on the training folds only, chosen via nested cross-validation, or tuned on all data? If the latter, the experiments must be rerun with a leakage-free protocol.
  2. [Section IV-B, Table IV] The abstract and Section V state that AcT with Integrated Gradients 'outperforms' and 'surpasses' the ground-truth-trained frame-level baseline (72% vs. 69%), but the legend of Table IV marks this comparison as not statistically significant (†, p>0.05). A non-significant point-estimate advantage does not support the wording 'better outcomes' or the conclusion that pseudo-labels are 'more robust than manual annotations.' Please report confidence intervals or effect sizes for these comparisons, and either soften the claims to describe a non-significant trend or provide additional evidence for superiority.
  3. [Section IV-A, Table III] The conclusions state that MOMENT 'outperforms' the task-specific models AcT and SkateFormer in video-level assessment, but Table III reports no statistically significant difference among the three pretrained models (p>0.05 for paired t-tests). The only significant comparisons are AcT vs. LSTM and MOMENT vs. LSTM. The abstract's 'MOMENT achieves better video-level assessment results' should either be explicitly restricted to the comparison with the LSTM baseline or supported by a proper multiple-comparison analysis that accounts for the three pairwise pretrained-model comparisons.
minor comments (8)
  1. [Section II-A] There is a typo in 'shoulder elevatio'; it should be 'shoulder elevation.'
  2. [Section III-A3] The sentence 'with 9.33±1.25 yeas of experience' contains a typo; 'yeas' should be 'years.'
  3. [Section IV-A] The phrase 'statiscally significant difference' should be corrected to 'statistically significant difference.'
  4. [Table IV] The Ground Truth column shows a single value (0.69 ± 0.11) shared across both threshold rows; please clarify that this is the same ground-truth-trained baseline compared against every configuration, and describe how the paired t-tests were computed across the 18 LOSO folds.
  5. [Section III-B] The statement 'Since E1 is performed with the left arm' appears to refer to a specific patient's affected side; clarify whether this is an example and whether the models and pseudo-labeling are invariant to left/right mirroring.
  6. [Section II-E, Eqs. (1)-(2)] The pseudo-label equations use the predicted class ŷ_i rather than the true video-level label, so errors made by Model A propagate directly into the pseudo-labels. The paper should explicitly discuss this error-propagation mechanism and its potential effect on Model B training.
  7. [Section II-C] The moving average window of five frames is introduced without justification; if alternative window sizes were considered in preliminary experiments, reporting them would help readers assess the sensitivity of the results.
  8. [References] Reference [20], which is the source of the SERE dataset, is a TechRxiv preprint; if a more stable journal or conference version exists, it should be cited.

Circularity Check

0 steps flagged · score 2.0 of 10

No load-bearing circularity; central weak-supervision pipeline is internally consistent, though threshold-selection protocol and significance claims are separate correctness concerns.

full rationale

Walking the claimed derivation chain: Model A is trained on video-level labels; gradient saliency is aggregated and thresholded (Eqs. 1-2) to form pseudo-labels; Model B is trained on these pseudo-labels and evaluated by AUC against held-out frame-level ground truth. No equation in the paper makes the evaluation labels an input to pseudo-label construction. The only potentially circular element is the statement in Section III-B that threshold values (Table II) were determined for each model to balance the ratio between False Positives and False Negatives. Taken literally, balancing FP/FN requires knowing which pseudo-labels are wrong, i.e., frame-level truth; if that balancing used held-out labels, the frame-level AUC comparison would be fitted rather than predicted. However, the paper also states that frame-level ground-truth labels are only used for training the baseline frame-level classifier and for validation purposes (Section I), and it does not describe any protocol that fits thresholds to validation labels. Under hard rule 1, I cannot exhibit the specific reduction; this is a transparency/correctness risk, not an established circular step. Self-citations to [20] (method and SERE dataset) and [30] (MOMENT) are real but not load-bearing: the pseudo-label method is fully specified in the present text, MOMENT is an independently pre-trained public model, and the empirical evaluation is LOSO on held-out patients. The headline claim that AcT-IG outperforms the ground-truth-trained baseline is statistically unsupported (Table IV legend: p > 0.05 vs Ground Truth), but non-significance is a correctness issue, not circularity. Score 2 reflects minor non-load-bearing self-citation without any derivation step that reduces by construction to its own inputs.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

No new physical entities are introduced. The main tunable components are the pseudo-label thresholds and MLP hyperparameters, both fitted to the SERE data; the preprocessing choices add additional hand-selected constants.

free parameters (3)
  • Pseudo-label thresholds (tau, tau1, tau2) = Table II: e.g., AcT IG single 2.7, dual 2.2/3.2
    Tuned per model and gradient type to balance false positive and false negative ratios; directly determines frame pseudo-labels, Section III-B.
  • MLP architecture and training hyperparameters = 1-2 layers, hidden units 32-256, LR 0.001/0.0001/0.00001, dropout 0.2/0.3, batch 16/32
    Grid searched for generalization; affects frame-level AUC, Section IV.
  • Moving average window = 5 frames
    Hand-chosen smoothing; no ablation reported, Section II-C.
assumptions (4)
  • domain assumption Video-level labels y_i are reliable and indicate presence of compensation somewhere in the trial.
    Used as ground truth for Model A and to gate pseudo-label generation, Eq. (1)-(2).
  • domain assumption Input-gradient saliency magnitude is monotonically related to frame-level presence of compensation.
    Core assumption of pseudo-label selection, Section II-E.
  • domain assumption MediaPipe 3D keypoints capture sufficient information for compensation detection.
    Pipeline input; no clinical or appearance features, Section II-C.
  • domain assumption Transfer from action-recognition or time-series pretraining to stroke pose data is beneficial.
    Motivates choice of AcT, SkateFormer, MOMENT, Section II-D.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Frame-Level Real-Time Assessment of Stroke Rehabilitation Exercises from Video-Level Labeled Data: Task-Specific vs. Foundation Models." pith.science (2026). https://pith.science/paper/4UKTYOHH

@misc{pith2026250603752,
  author       = {Pith},
  title        = {Pith review of: Frame-Level Real-Time Assessment of Stroke Rehabilitation Exercises from Video-Level Labeled Data: Task-Specific vs. Foundation Models},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/4UKTYOHH}},
  note         = {Machine review of arXiv:2506.03752}
}
abstract

The growing demands of stroke rehabilitation have increased the need for solutions to support autonomous exercising. Virtual coaches can provide real-time exercise feedback from video data, helping patients improve motor function and keep engagement. However, training real-time motion analysis systems demands frame-level annotations, which are time-consuming and costly to obtain. In this work, we present a framework that learns to classify individual frames from video-level annotations for real-time assessment of compensatory motions in rehabilitation exercises. We use a gradient-based technique and a pseudo-label selection method to create frame-level pseudo-labels for training a frame-level classifier. We leverage pre-trained task-specific models - Action Transformer, SkateFormer - and a foundation model - MOMENT - for pseudo-label generation, aiming to improve generalization to new patients. To validate the approach, we use the \textit{SERE} dataset with 18 post-stroke patients performing five rehabilitation exercises annotated on compensatory motions. MOMENT achieves better video-level assessment results (AUC = $73\%$), outperforming the baseline LSTM (AUC = $58\%$). The Action Transformer, with the Integrated Gradient technique, leads to better outcomes (AUC = $72\%$) for frame-level assessment, outperforming the baseline trained with ground truth frame-level labeling (AUC = $69\%$). We show that our proposed approach with pre-trained models enhances model generalization ability and facilitates the customization to new patients, reducing the demands of data labeling.

Figures

Figures reproduced from arXiv: 2506.03752 by the authors.

Figure 2
Figure 2. Approach Pipeline: a) body pose extraction and [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗
Figure 3
Figure 3. MediaPipe Pose Keypoints. The pose estimation and pre￾processing steps extract key motion features, reduce noise, and enhance model general￾ization. MediaPipe1 is used to track post-stroke patients’ movements, providing 33 joint keypoints per frame in real￾world 3D coordinates [PITH_FULL_IMAGE:figures/full_fig_p003_3.png] view at source ↗
Figure 5
Figure 5. Saliency map generated by the AcT model using the [PITH_FULL_IMAGE:figures/full_fig_p004_5.png] view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: SERE functional exercises for rehabilitation. saliency map, these joints are highlighted more prominently than in the case of a normal (non-compensatory) motion, indicating a compensatory behavior. Additionally, for the pseudo-label creation, we inspect the benefit of …

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

39 extracted references · 31 canonical work pages

  1. [1]

    Task-oriented training in rehabilitation after stroke: systematic re- view,

    M. Rensink, M. Schuurmans, E. Lindeman, and T. Hafsteinsdottir, “Task-oriented training in rehabilitation after stroke: systematic re- view,”Journal of advanced nursing, vol. 65, no. 4, pp. 737–754, 2009

    work page 2009

  2. [2]

    Extra upper limb practice after stroke: a feasibility study,

    E. J. Schneider, L. Ada, and N. A. Lannin, “Extra upper limb practice after stroke: a feasibility study,”Pilot and Feasibility Studies, vol. 5, pp. 1–7, 2019

    work page 2019

  3. [3]

    Physical activity and exercise recommendations for stroke survivors: a statement for healthcare professionals from the american heart association/american stroke association,

    S. A. Billinger, R. Arena, J. Bernhardt, J. J. Eng, B. A. Franklin, C. M. Johnson, M. MacKay-Lyons, R. F. Macko, G. E. Mead, E. J. Rothet al., “Physical activity and exercise recommendations for stroke survivors: a statement for healthcare professionals from the american heart association/american stroke association,”Stroke, vol. 45, no. 8, pp. 2532–2553, 2014

    work page 2014

  4. [4]

    What is current practice for upper limb rehabilitation in the acute hospital setting following stroke? a systematic review,

    I. Serrada, M. N. McDonnell, and S. L. Hillier, “What is current practice for upper limb rehabilitation in the acute hospital setting following stroke? a systematic review,”NeuroRehabilitation, vol. 39, no. 3, pp. 431–438, 2016

    work page 2016

  5. [5]

    Factors affecting rehabilitation and use of upper limb after stroke: views from healthcare professionals and stroke survivors,

    K. L. Meadmore, E. Hallewell, C. Freeman, and A.-M. Hughes, “Factors affecting rehabilitation and use of upper limb after stroke: views from healthcare professionals and stroke survivors,”Topics in stroke rehabilitation, vol. 26, no. 2, pp. 94–100, 2019

    work page 2019

  6. [6]

    Challenges in applying evidence-based practice in stroke rehabilita- tion: a qualitative description of health professional experience in low, middle, and high-income countries,

    K. E. Watkins, W. M. Levack, F. A. Rathore, and E. J. C. Hay-Smith, “Challenges in applying evidence-based practice in stroke rehabilita- tion: a qualitative description of health professional experience in low, middle, and high-income countries,”Disability and Rehabilitation, pp. 1–9, 2023

    work page 2023

  7. [7]

    Interventions to aid patient adherence to physiotherapist prescribed self-management strategies: a systematic review,

    K. Peek, R. Sanson-Fisher, L. Mackenzie, and M. Carey, “Interventions to aid patient adherence to physiotherapist prescribed self-management strategies: a systematic review,”Physiotherapy, vol. 102, no. 2, pp. 127–135, 2016

    work page 2016

  8. [8]

    Qualitative analysis of stroke patients’ motivation for rehabilitation,

    N. Maclean, P. Pound, C. Wolfe, and A. Rudd, “Qualitative analysis of stroke patients’ motivation for rehabilitation,”Bmj, vol. 321, no. 7268, pp. 1051–1054, 2000

    work page 2000

Show all 39 references
  1. [9]

    Architecture and applica- tions of virtual coaches,

    D. Siewiorek, A. Smailagic, and A. Dey, “Architecture and applica- tions of virtual coaches,”Proceedings of the IEEE, vol. 100, no. 8, pp. 2472–2488, 2012

    2012

  2. [10]

    Virtual coaches: background, theories, and future research directions,

    T. G. Weimann, H. Schlieter, and A. B. Brendel, “Virtual coaches: background, theories, and future research directions,”Business & Information Systems Engineering, vol. 64, no. 4, pp. 515–528, 2022

    2022

  3. [11]

    A clinically feasible kinematic assessment method of upper extremity motor function impairment after stroke,

    A. Ozturk, A. Tartar, B. E. Huseyinsinoglu, and A. H. Ertas, “A clinically feasible kinematic assessment method of upper extremity motor function impairment after stroke,”Measurement, vol. 80, pp. 207–216, 2016

    2016

  4. [12]

    Automated assessment of upper extremity movement impairment due to stroke,

    E. V . Olesh, S. Yakovenko, and V . Gritsenko, “Automated assessment of upper extremity movement impairment due to stroke,”PloS one, vol. 9, no. 8, p. e104487, 2014

    2014

  5. [13]

    Learning to assess the quality of stroke rehabilitation exercises,

    M. H. Lee, D. P. Siewiorek, A. Smailagic, A. Bernardino, and S. B. i. Badia, “Learning to assess the quality of stroke rehabilitation exercises,” inInternational Conference on Intelligent User Interfaces, 2019, pp. 218–228

    2019

  6. [14]

    Ai in rehabilitation medicine: Opportunities and challenges,

    F. Lanotte, M. K. O’Brien, and A. Jayaraman, “Ai in rehabilitation medicine: Opportunities and challenges,”Annals of Rehabilitation Medicine, vol. 47, no. 6, p. 444, 2023

    2023

  7. [15]

    Learning from few examples: A sum- mary of approaches to few-shot learning. arxiv 2022,

    A. Parnami and M. Lee, “Learning from few examples: A sum- mary of approaches to few-shot learning. arxiv 2022,”arXiv preprint arXiv:2203.04291, 2022

    2022 arXiv

  8. [16]

    Towards personalized interaction and corrective feedback of a socially assistive robot for post-stroke rehabilitation therapy,

    M. H. Lee, D. P. Siewiorek, A. Smailagic, A. Bernardino, and S. B. Badia, “Towards personalized interaction and corrective feedback of a socially assistive robot for post-stroke rehabilitation therapy,” in IEEE International Conference on Robot and Human Interactive Communicat...

    2020

  9. [17]

    A low-cost virtual coach for 2d video-based compensation assessment of upper extremity reha- bilitation exercises,

    A. R. C ´oias, M. H. Lee, and A. Bernardino, “A low-cost virtual coach for 2d video-based compensation assessment of upper extremity reha- bilitation exercises,”Journal of NeuroEngineering and Rehabilitation, vol. 19, no. 1, pp. 1–16, 2022

    2022

  10. [18]

    A deep learning system to monitor and assess rehabilitation exercises in home- based remote and unsupervised conditions,

    C. Mennella, U. Maniscalco, G. De Pietro, and M. Esposito, “A deep learning system to monitor and assess rehabilitation exercises in home- based remote and unsupervised conditions,”Computers in Biology and Medicine, vol. 166, p. 107485, 2023

    2023

  11. [19]

    Exploring a gradient-based explainable ai technique for time-series data: A case study of assessing stroke rehabilitation exercises,

    M. H. Lee and Y . J. Choy, “Exploring a gradient-based explainable ai technique for time-series data: A case study of assessing stroke rehabilitation exercises,”arXiv preprint arXiv:2305.05525, 2023

    2023 arXiv

  12. [20]

    Learning frame-level classifiers for video-based real-time assessment of stroke rehabilitation exercises from weakly annotated datasets,

    A. R. C ´oias, M. H. Lee, A. Bernardino, A. Smailagic, M. Mateus, D. Fernandes, and S. Trapola, “Learning frame-level classifiers for video-based real-time assessment of stroke rehabilitation exercises from weakly annotated datasets,”TechRxiv, Jan. 2025

    2025

  13. [21]

    Segment anything model (sam) enhanced pseudo labels for weakly supervised semantic seg- mentation,

    T. Chen, Z. Mai, R. Li, and W.-l. Chao, “Segment anything model (sam) enhanced pseudo labels for weakly supervised semantic seg- mentation,”arXiv preprint arXiv:2305.05803, 2023

    2023 arXiv

  14. [22]

    Frame-level label refinement for skeleton- based weakly-supervised action recognition,

    Q. Yu and K. Fujiwara, “Frame-level label refinement for skeleton- based weakly-supervised action recognition,” inProceedings of the AAAI Conference on Artificial Intelligence, vol. 37, no. 3, 2023, pp. 3322–3330

    2023

  15. [23]

    Cross-model pseudo-labeling for semi-supervised action recognition,

    Y . Xu, F. Wei, X. Sun, C. Yang, Y . Shen, B. Dai, B. Zhou, and S. Lin, “Cross-model pseudo-labeling for semi-supervised action recognition,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022, pp. 2959–2968

    2022

  16. [24]

    Transfer learning: a friendly introduction,

    A. Hosna, E. Merry, J. Gyalmo, Z. Alom, Z. Aung, and M. A. Azim, “Transfer learning: a friendly introduction,”Journal of Big Data, vol. 9, no. 1, p. 102, 2022

    2022

  17. [25]

    Human action recognition based on transfer learning approach,

    Y . Abdulazeem, H. M. Balaha, W. M. Bahgat, and M. Badawy, “Human action recognition based on transfer learning approach,”IEEE Access, vol. 9, pp. 82 058–82 069, 2021

    2021

  18. [26]

    Internal trans- fer learning for improving performance in human action recognition for small datasets,

    T. Wang, Y . Chen, M. Zhang, J. Chen, and H. Snoussi, “Internal trans- fer learning for improving performance in human action recognition for small datasets,”IEEE Access, vol. 5, pp. 17 627–17 633, 2017

    2017

  19. [27]

    Skelresnet: Transfer learning approach for skeleton-based action recognition,

    U. Kılıc ¸, ¨O. ¨O. Karada ˘g, and G. T. ¨Ozyer, “Skelresnet: Transfer learning approach for skeleton-based action recognition,” in2024 32nd Signal Processing and Communications Applications Conference (SIU). IEEE, 2024, pp. 1–4

    2024

  20. [28]

    Action transformer: A self-attention model for short-time pose-based human action recognition,

    V . Mazzia, S. Angarano, F. Salvetti, F. Angelini, and M. Chiaberge, “Action transformer: A self-attention model for short-time pose-based human action recognition,”Pattern Recognition, vol. 124, p. 108487, 2022

    2022

  21. [29]

    Skateformer: skeletal-temporal transformer for human action recognition,

    J. Do and M. Kim, “Skateformer: skeletal-temporal transformer for human action recognition,” inEuropean Conference on Computer Vision. Springer, 2025, pp. 401–420

    2025

  22. [30]

    Moment: A family of open time-series foundation models,

    M. Goswami, K. Szafer, A. Choudhry, Y . Cai, S. Li, and A. Dubrawski, “Moment: A family of open time-series foundation models,”arXiv preprint arXiv:2402.03885, 2024

    2024 arXiv

  23. [31]

    What do motor “recovery

    M. F. Levin, J. A. Kleim, and S. L. Wolf, “What do motor “recovery” and “compensation” mean in patients following stroke?”Neuroreha- bilitation and neural repair, vol. 23, no. 4, pp. 313–319, 2009

    2009

  24. [33]

    Axiomatic attribution for deep networks,

    M. Sundararajan, A. Taly, and Q. Yan, “Axiomatic attribution for deep networks,” inInternational conference on machine learning. PMLR, 2017, pp. 3319–3328

    2017

  25. [34]

    Skeleton tracking solutions for a low-cost stroke rehabilitation support system,

    A. R. C ´oias, M. H. Lee, A. Bernardino, and A. Smailagic, “Skeleton tracking solutions for a low-cost stroke rehabilitation support system,” in2023 International Conference on Rehabilitation Robotics (ICORR). IEEE, 2023, pp. 1–6

    2023

  26. [35]

    Dosovitskiy, L

    A. Dosovitskiy, L. Beyer, A. Kolesnikov, D. Weissenborn, and X. Zhai, “Thomas unterthiner mostafa dehghani matthias minderer georg heigold sylvain gelly jakob uszkoreit and neil houlsby. an image isworth 16×16 words: Transformers for image recognition atscale,” inInternational...

    2021

  27. [36]

    Ntu rgb+ d 120: A large-scale benchmark for 3d human activity understanding,

    J. Liu, A. Shahroudy, M. Perez, G. Wang, L.-Y . Duan, and A. C. Kot, “Ntu rgb+ d 120: A large-scale benchmark for 3d human activity understanding,”IEEE transactions on pattern analysis and machine intelligence, vol. 42, no. 10, pp. 2684–2701, 2019

    2019

  28. [37]

    Ntu rgb+ d: A large scale dataset for 3d human activity analysis,

    A. Shahroudy, J. Liu, T.-T. Ng, and G. Wang, “Ntu rgb+ d: A large scale dataset for 3d human activity analysis,” inProceedings of the IEEE conference on computer vision and pattern recognition, 2016, pp. 1010–1019

    2016

  29. [38]

    Cross-view action modeling, learning and recognition,

    J. Wang, X. Nie, Y . Xia, Y . Wu, and S.-C. Zhu, “Cross-view action modeling, learning and recognition,” inProceedings of the IEEE conference on computer vision and pattern recognition, 2014, pp. 2649–2656

    2014

  30. [39]

    Deep Inside Convolutional Networks: Visualising Image Classification Models and Saliency Maps,

    K. Simonyan, A. Vedaldi, and A. Zisserman, “Deep Inside Convolutional Networks: Visualising Image Classification Models and Saliency Maps,” Apr. 2014, arXiv:1312.6034 [cs]. [Online]. Available: http://arxiv.org/abs/1312.6034

    2014 arXiv

  31. [40]

    Axiomatic Attribution for Deep Networks,

    M. Sundararajan, A. Taly, and Q. Yan, “Axiomatic Attribution for Deep Networks,” Jun. 2017, arXiv:1703.01365 [cs]. [Online]. Available: http://arxiv.org/abs/1703.01365

    2017 arXiv

Pith tools

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