pith. sign in

arxiv: 2606.28104 · v1 · pith:5HOBLMYOnew · submitted 2026-06-26 · 💻 cs.CV · cs.LG

Cross-view Multimodal Vision-Based Assessment Framework for Traditional Chinese Medicine Rehabilitation Training

Francis Xiatian Zhang , Hao Yao , Shengxuan Chen , Hong Zhu , Hongxiao Jia , Sisi Zheng , Hubert P. H. Shum This is my paper

Pith reviewed 2026-06-29 04:00 UTC · model grok-4.3

classification 💻 cs.CV cs.LG
keywords action quality assessmenttraditional chinese medicinecross-view videomultimodal fusionacupuncture assessmenttuina assessmentvision-based evaluationrehabilitation training
0
0 comments X

The pith

A cross-view video system with visual-pose fusion improves automated scoring of acupuncture and Tuina skills.

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

The paper presents CME-AQA as a way to assess Traditional Chinese Medicine rehabilitation actions from video. It claims single-view methods fail on hand-heavy tasks because of occlusion and interactions, so combining first-person and third-person footage plus pose data gives better context. New dual-view datasets of acupuncture and Tuina performed by 61 subjects with expert labels serve as the test bed. Results show gains over baselines on rating tasks and lower error on timing and frequency measures. The same setup also matches performance on a CPR posture dataset, pointing to wider use in structured clinical skill checks.

Core claim

The central claim is that the CME-AQA framework, which fuses visual and pose information from synchronized first-person and third-person videos, delivers superior or comparable mean performance to competitive baselines on action quality assessment for TCM techniques, including more than 10 percent relative improvement in weighted F1 on Needle Depth and Quick Needle Insertion tasks along with reduced mean absolute error for insertion time and manipulation frequency, while producing comparable results on several posture criteria in a CPR dataset.

What carries the argument

The CME-AQA framework that performs cross-view multimodal fusion of visual and pose data from first-person and third-person videos to handle self-occlusion and hand-object interactions.

If this is right

  • Rating accuracy rises for TCM tasks that involve dense hand movements and object interactions.
  • Quantitative measures such as insertion time and manipulation frequency are estimated with lower error than single-view baselines.
  • The approach transfers to other structured clinical skill assessments centered on participant posture and motion, such as CPR.
  • Video-based evaluation reduces reliance on constant in-person expert observation during training.

Where Pith is reading between the lines

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

  • The system could support remote or self-paced feedback loops for TCM students without requiring an instructor to be physically present.
  • Real deployment would need checks against broader ranges of patient body types and practitioner experience levels beyond the 61-subject sets.
  • Adding non-visual signals like force sensors on needles could strengthen the multimodal component for finer manipulation scoring.

Load-bearing premise

The collected dual-view videos with expert annotations on 61 subjects represent the variability present in actual clinical TCM training sessions.

What would settle it

Performance gains disappear when the same model is tested on videos from a new group of practitioners and patients recorded under different lighting or camera placements not seen in the original TCM-AQA61 collections.

Figures

Figures reproduced from arXiv: 2606.28104 by Francis Xiatian Zhang, Hao Yao, Hongxiao Jia, Hong Zhu, Hubert P. H. Shum, Shengxuan Chen, Sisi Zheng.

Figure 1
Figure 1. Figure 1: Example frames from our two collected datasets: acupuncture (upper rows) and Tuina (lower rows). Each session was recorded with two cameras: a forehead-mounted first-person view (left column) and a separate third-person view of the hands (right column), providing complementary participant and observer perspectives. I. INTRODUCTION T Raditional Chinese Medicine (TCM) rehabilitation ther￾apies, such as acupu… view at source ↗
Figure 2
Figure 2. Figure 2: Dataset statistics and label distribution for TCM-AQA61-A (Acupuncture) and TCM-AQA61-T (Tuina). (a)–(b) Binary distributions of negative (0) and positive (1) annotations for each assessment indicator across 61 subjects. Although some indicators exhibit moderate class im￾balance (e.g., Twisting Frequency in Acupuncture and Slow Movement in Tuina), no task collapses to a single dominant class. (c) Distribut… view at source ↗
Figure 3
Figure 3. Figure 3: Overview of the Cross-view Multimodality Enhanced Action Quality Assessment (CME-AQA) framework, which leverages multi-view and multimodal data for TCM rehabilitation assessment. It integrates two main components: (1) the Attention-based Visual–Pose Fusion (AVPF) module (Section IV-A), which fuses visual and pose features through cross- and self-attention, and (2) the Multiscale View Alignment (MVA) traini… view at source ↗
Figure 4
Figure 4. Figure 4: Architecture of the cross-attention transformer in our AVPF mod￾ule. Visual features (Q) are refined with pose features (K, V ), while pose features remain fixed. This design enables the model to attend to pose cues relevant to the visual context, supporting interaction-aware feature fusion for AQA [PITH_FULL_IMAGE:figures/full_fig_p006_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: The architecture of our layer-normalization-enhanced attention transformer block. To mitigate potential noise caused by pose estimation inaccuracies, an additional normalization layer is applied after the 1D convolution to enhance robustness. stable pose structure. For implementation, visual features are extracted using a pretrained backbone (e.g., ResNet [51]) to obtain FV ∈ R T ×C . Hand pose features ar… view at source ↗
Figure 7
Figure 7. Figure 7: Illustration of complementary motion cues captured by third￾person (TPV) and first-person (FPV) views in TCM-AQA61-A during needle insertion. (a) Well-performed trial where the needle is inserted stably and quickly; (b) poorly performed trial where the needle is inserted irregularly and slowly. In each case, the upper row shows TPV frames and the lower row shows FPV frames. TPV better exposes the vertical … view at source ↗
Figure 6
Figure 6. Figure 6: Architecture of the self-attention transformer in our AVPF mod￾ule. Fused features act as Q, K, and V , allowing the model to refine representations by focusing on the most relevant information for AQA inference. F I V , while pose features remain unchanged throughout. The fused representation F I V is further processed by a stack of layer-normalized temporal self-attention blocks ( [PITH_FULL_IMAGE:figur… view at source ↗
Figure 8
Figure 8. Figure 8: Example frames of the views used. For fair comparison, both the front and side views were included during training, as these were the main viewpoints used by experts, while only the front view was used for inference. on an NVIDIA GeForce RTX 4090 GPU. For training and testing, we used a fivefold cross-validation protocol, allocating 80% of the data for training and 20% for testing. Following Constable et a… view at source ↗
Figure 9
Figure 9. Figure 9: Critical case comparison for Quick Needle Insertion. The red box highlights the hand–needle interaction. The well-performing trial completes insertion within 3s, whereas the poorly performing trial lasts about 10s with slower motion. CME-AQA correctly differentiates the two cases, while strong baselines misclassify at least one instance. See supplementary videos for full sequences. E. Qualitative Analysis … view at source ↗
Figure 11
Figure 11. Figure 11: Illustration of progressive pose perturbation. From left to right: original pose (no additional noise), and injected Gaussian jitter with σ = 1σemp, 2σemp, 3σemp. Noise is added independently to each joint coordinate in normalized space. As σ increases, joint locations exhibit progressively larger spatial deviations, leading to increasing geometric distortion of finger alignment and hand contour while pre… view at source ↗
Figure 10
Figure 10. Figure 10: Temporal attention patterns learned by CME-AQA for the same well-performing (a) and poorly performing (b) trials. Shaded regions denote insertion (green) and withdrawal (brown) phases. In the well-performing case, attention is sharply concentrated within the brief insertion window, whereas in the poorly performing case it is temporally dispersed across the prolonged insertion period. This behavior indicat… view at source ↗
read the original abstract

Vision-based assessment can provide convenient and cost-effective evaluation in Traditional Chinese Medicine (TCM) rehabilitation training, where action quality assessment (AQA) from computer vision offers a promising solution. Existing automatic AQA frameworks for physical therapy typically rely on skeletal data captured from a single viewpoint, which is inefficient for TCM techniques such as acupuncture or Tuina that involve dense hand self-occlusion and complex hand-object interactions. To address these challenges, we propose CME-AQA, a cross-view, multimodal vision-based assessment framework that integrates visual-pose fusion to enhance understanding of environmental context and leverages both first-person and third-person videos during training to improve inference robustness. We collected two dual-view datasets, TCM-AQA61-A (Acupuncture) and TCM-AQA61-T (Tuina), each containing synchronized first-person and third-person recordings of 61 subjects with expert annotations. Experimental results show that our approach achieves superior or comparable mean performance against competitive baselines, achieving over 10% relative improvement in weighted F1 over the best competing method on key rating tasks such as Needle Depth and Quick Needle Insertion, while also reducing mean absolute error in quantitative measures such as insertion time and manipulation frequency. Testing on a CPR dataset further demonstrates comparable performance on several posture-based criteria, suggesting applicability to related structured simulated clinical skill assessments where participant motion is central to evaluation. Overall, CME-AQA enhances assessment accuracy for structured TCM rehabilitation training and facilitates more convenient and effective training-oriented skill evaluation.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 2 minor

Summary. The manuscript introduces CME-AQA, a cross-view multimodal vision-based framework for action quality assessment (AQA) in Traditional Chinese Medicine rehabilitation training (acupuncture and Tuina). It combines visual-pose fusion with synchronized first-person and third-person videos to handle hand self-occlusion and hand-object interactions, unlike single-view skeletal approaches. The authors collect two new dual-view datasets (TCM-AQA61-A and TCM-AQA61-T) from 61 subjects with expert annotations, report >10% relative weighted-F1 gains on tasks such as Needle Depth and Quick Needle Insertion versus baselines, reduced MAE on quantitative measures like insertion time, and comparable results when tested on a CPR dataset.

Significance. If the performance deltas are shown to arise from the proposed cross-view fusion rather than dataset or input-format advantages, the work would strengthen multi-view video methods for assessing complex, occluded motor skills in clinical training. The new TCM-specific dual-view datasets and the CPR transfer experiment provide concrete resources and evidence of broader applicability. These elements would be notable strengths for a computer-vision journal if the experimental controls are tightened.

major comments (2)
  1. [Abstract and §4] Abstract and §4 (Experiments): the central claim of >10% relative weighted-F1 improvement on Needle Depth and Quick Needle Insertion rests on comparisons whose fairness cannot be verified from the given information; no description is supplied of how single-view baselines were adapted (or not) to consume the same synchronized first-/third-person pairs that CME-AQA receives at training and test time.
  2. [§3] §3 (Dataset): the weakest assumption—that expert annotations on TCM-AQA61 constitute unbiased, reproducible ground truth that generalizes to clinical variability—is unsupported; the manuscript supplies no inter-rater reliability statistics, annotation protocol, or analysis of subject or action diversity.
minor comments (2)
  1. [Abstract] The abstract states that the method 'leverages both first-person and third-person videos during training to improve inference robustness,' yet the precise training/inference split and whether third-person views are available at test time are not stated.
  2. Notation for the visual-pose fusion weights and backbone hyperparameters is introduced without an accompanying equation or table reference.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive feedback. We address each major comment below and indicate planned revisions.

read point-by-point responses

  1. Referee: [Abstract and §4] Abstract and §4 (Experiments): the central claim of >10% relative weighted-F1 improvement on Needle Depth and Quick Needle Insertion rests on comparisons whose fairness cannot be verified from the given information; no description is supplied of how single-view baselines were adapted (or not) to consume the same synchronized first-/third-person pairs that CME-AQA receives at training and test time.

    Authors: We agree that the input configurations require explicit clarification to isolate the contribution of cross-view fusion. The single-view baselines were evaluated on the third-person view alone, consistent with their original design, while CME-AQA receives synchronized dual-view pairs. In revision we will expand §4 with a precise description of each method's inputs and add an ablation comparing our model under single-view versus dual-view conditions to demonstrate that gains arise from the proposed fusion rather than input format. revision: yes

  2. Referee: [§3] §3 (Dataset): the weakest assumption—that expert annotations on TCM-AQA61 constitute unbiased, reproducible ground truth that generalizes to clinical variability—is unsupported; the manuscript supplies no inter-rater reliability statistics, annotation protocol, or analysis of subject or action diversity.

    Authors: We acknowledge the need for greater transparency on annotation quality. The revised §3 will include the annotation protocol (conducted by certified TCM experts using standardized rubrics) and a summary table of subject demographics and action distributions. Inter-rater reliability statistics were not collected during dataset creation; we will explicitly state this limitation and its implications for generalizability. revision: partial

Circularity Check

0 steps flagged

No significant circularity; empirical evaluation on held-out expert-annotated data is independent of model construction.

full rationale

The paper proposes CME-AQA, a cross-view multimodal framework, and reports empirical performance gains on newly collected TCM-AQA61 datasets with expert annotations. No equations or claims reduce a prediction to a fitted input by construction, no self-citation chain bears the central result, and no uniqueness theorem or ansatz is smuggled in. The performance numbers are measured on separate test splits against adapted baselines; the derivation chain (feature fusion and training procedure) remains self-contained and externally falsifiable via the reported metrics.

Axiom & Free-Parameter Ledger

2 free parameters · 1 axioms · 0 invented entities

The central claim rests on the quality of expert annotations, the representativeness of the 61-subject dual-view recordings, and the effectiveness of an unspecified visual-pose fusion architecture whose parameters are learned from data.

free parameters (2)
  • visual-pose fusion weights
    Parameters that combine visual and pose streams are learned from the training split and directly affect reported F1 and MAE numbers.
  • backbone hyperparameters
    Standard deep network hyperparameters tuned on the collected data.
axioms (1)
  • domain assumption Expert annotations constitute reliable ground truth for action quality
    All reported metrics depend on these labels being accurate and consistent across raters.

pith-pipeline@v0.9.1-grok · 5818 in / 1189 out tokens · 59628 ms · 2026-06-29T04:00:28.838596+00:00 · methodology

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

64 extracted references · 3 canonical work pages · 1 internal anchor

  1. [1]

    W.-W. Tao, H. Jiang, X.-M. Tao, P. Jiang, L.-Y . Sha, and X.-C. Sun, “Effects of acupuncture, tuina, tai chi, qigong, and traditional chinese medicine five-element music therapy on symptom management and quality of life for cancer patients: a meta-analysis,”Journal of pain and symptom management, vol. 51, no. 4, pp. 728–747, 2016. 1

    2016

  2. [2]

    Effect of acupuncture treatment on cortical activation in patients with tinnitus: a functional near-infrared spectroscopy study,

    X. Yu, B. Gong, H. Yang, Z. Wang, G. Qi, J. Sun, Y . Fang, and X. Fan, “Effect of acupuncture treatment on cortical activation in patients with tinnitus: a functional near-infrared spectroscopy study,” IEEE Transactions on Neural Systems and Rehabilitation Engineering, vol. 31, pp. 729–737, 2023. 1

    2023

  3. [3]

    Evidence-based practice of chinese medicine in physical rehabilitation science,

    A. de S ´a Ferreira, “Evidence-based practice of chinese medicine in physical rehabilitation science,”Chinese journal of integrative medicine, vol. 19, no. 10, pp. 723–729, 2013. 1

    2013

  4. [4]

    A mobile natural human–robot interaction method for virtual chinese acupuncture,

    G. Du, Y . Li, K. Su, C. Li, and P. X. Liu, “A mobile natural human–robot interaction method for virtual chinese acupuncture,”IEEE Transactions on Instrumentation and Measurement, vol. 72, pp. 1–10, 2022. 1, 3, 5

    2022

  5. [5]

    Design and development of a mixed reality acupuncture training system,

    Q. Sun, J. Huang, H. Zhang, P. Craig, L. Yu, and E. G. Lim, “Design and development of a mixed reality acupuncture training system,” in2023 IEEE Conference Virtual Reality and 3D User Interfaces (VR). IEEE, 2023, pp. 265–275. 1

    2023

  6. [6]

    Ai-driven stroke rehabilitation systems and assessment: a systematic review,

    S. Rahman, S. Sarker, A. N. Haque, M. M. Uttsha, M. F. Islam, and S. Deb, “Ai-driven stroke rehabilitation systems and assessment: a systematic review,”IEEE Transactions on Neural Systems and Reha- bilitation Engineering, vol. 31, pp. 192–207, 2022. 2

    2022

  7. [7]

    Vision-based human action quality assessment: A systematic review,

    J. Liu, H. Wang, K. Stawarz, S. Li, Y . Fu, and H. Liu, “Vision-based human action quality assessment: A systematic review,”Expert Systems with Applications, p. 125642, 2024. 2

    2024

  8. [8]

    A survey on video action recognition in sports: Datasets, methods and applications,

    F. Wu, Q. Wang, J. Bian, N. Ding, F. Lu, J. Cheng, D. Dou, and H. Xiong, “A survey on video action recognition in sports: Datasets, methods and applications,”IEEE Transactions on Multimedia, 2022. 2

    2022

  9. [9]

    Standards for reporting interventions in clinical trials of tuina/massage (strictotm): Extending the consort statement,

    X. Zhang, F. Liang, C. T. Lau, J. C. Chan, N. Wang, J. Deng, J. Wang, Y . Ma, L. L. Zhong, C. Zhaoet al., “Standards for reporting interventions in clinical trials of tuina/massage (strictotm): Extending the consort statement,”Journal of Evidence-Based Medicine, vol. 16, no. 1, pp. 68–81, 2023. 2

    2023

  10. [10]

    A deep learning framework for assessing physical rehabilitation exercises,

    Y . Liao, A. Vakanski, and M. Xian, “A deep learning framework for assessing physical rehabilitation exercises,”IEEE Transactions on Neural Systems and Rehabilitation Engineering, vol. 28, no. 2, pp. 468– 477, 2020. 2, 3, 5, 6, 7, 8, 9

    2020

  11. [11]

    Graph convolutional networks for assessment of physical rehabilitation exercises,

    S. Deb, M. F. Islam, S. Rahman, and S. Rahman, “Graph convolutional networks for assessment of physical rehabilitation exercises,”IEEE Transactions on Neural Systems and Rehabilitation Engineering, vol. 30, pp. 410–419, 2022. 2, 8, 9 14 IEEE TRANSACTIONS AND JOURNALS TEMPLATE

    2022

  12. [12]

    A skeleton-based rehabilitation exercise assessment system with rotation invariance,

    K. Zheng, J. Wu, J. Zhang, and C. Guo, “A skeleton-based rehabilitation exercise assessment system with rotation invariance,”IEEE Transactions on Neural Systems and Rehabilitation Engineering, 2023. 2, 3, 5, 6, 8, 9

    2023

  13. [13]

    Dynamic multiview refinement of 3d hand datasets using differentiable ray tracing,

    G. Karvounas, N. Kyriazis, I. Oikonomidis, and A. Argyros, “Dynamic multiview refinement of 3d hand datasets using differentiable ray tracing,” inProceedings of the IEEE/CVF International Conference on Computer Vision, 2023, pp. 3156–3166. 2

    2023

  14. [14]

    Trends in acupuncture training research: focus on practical phantom models,

    J. E. Jang, Y . S. Lee, W. S. Jang, W. S. Sung, E.-J. Kim, S. D. Lee, K. H. Kim, and C. Y . Jung, “Trends in acupuncture training research: focus on practical phantom models,” 2022. 2, 3

    2022

  15. [15]

    A contrastive learning network for performance metric and assessment of physical rehabilita- tion exercises,

    L. Yao, Q. Lei, H. Zhang, J. Du, and S. Gao, “A contrastive learning network for performance metric and assessment of physical rehabilita- tion exercises,”IEEE Transactions on Neural Systems and Rehabilitation Engineering, 2023. 2, 3, 5, 6

    2023

  16. [16]

    Finediving: A fine-grained dataset for procedure-aware action quality assessment,

    J. Xu, Y . Rao, X. Yu, G. Chen, J. Zhou, and J. Lu, “Finediving: A fine-grained dataset for procedure-aware action quality assessment,” inProceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2022, pp. 2949–2958. 2

    2022

  17. [17]

    Logo: A long-form video dataset for group action quality assessment,

    S. Zhang, W. Dai, S. Wang, X. Shen, J. Lu, J. Zhou, and Y . Tang, “Logo: A long-form video dataset for group action quality assessment,” inProceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2023, pp. 2405–2414. 2

    2023

  18. [18]

    Pose- guided transformer for fine-grained action quality assessment,

    Y . Zhang, X. Li, W. Chai, C. Yan, W. Wang, and G. Wang, “Pose- guided transformer for fine-grained action quality assessment,”IEEE Transactions on Circuits and Systems for Video Technology, 2025. 2, 3, 5, 8, 9

    2025

  19. [19]

    Lucidaction: A hierarchical and multi-model dataset for comprehensive action quality assessment,

    L. Dong, W. Wang, Y . Qiao, and X. Sun, “Lucidaction: A hierarchical and multi-model dataset for comprehensive action quality assessment,” Advances in neural information processing systems, vol. 37, pp. 96 468– 96 482, 2024. 2, 3, 7

    2024

  20. [20]

    M. D. Constable, F. X. Zhang, T. Conner, D. Monk, J. Rajsic, C. Ford, L. J. Park, A. Platt, D. Porteous, L. Griersonet al., “Advancing healthcare practice and education via data sharing: demonstrating the utility of open data by training an artificial intelligence model to assess cardiopulmonary resuscitation skills,”Advances in Health Sciences Education,...

    2024

  21. [21]

    Clinical video analysis with geometric feature enhanced deep learning,

    X. Zhang, “Clinical video analysis with geometric feature enhanced deep learning,” Doctoral Thesis, Durham University, Durham, UK, 2025, available under Creative Commons Attribution 3.0 (CC BY). [Online]. Available: https://etheses.dur.ac.uk/16017/ 2

    2025

  22. [22]

    A survey of video-based action quality assessment,

    S. Wang, D. Yang, P. Zhai, Q. Yu, T. Suo, Z. Sun, K. Li, and L. Zhang, “A survey of video-based action quality assessment,” in2021 International conference on networking systems of AI (INSAI). IEEE, 2021, pp. 1–9. 3, 7

    2021

  23. [23]

    Assessing the quality of actions,

    H. Pirsiavash, C. V ondrick, and A. Torralba, “Assessing the quality of actions,” inEuropean conference on computer vision. Springer, 2014, pp. 556–571. 3

    2014

  24. [24]

    What and how well you performed? a mul- titask learning approach to action quality assessment,

    P. Parmar and B. T. Morris, “What and how well you performed? a mul- titask learning approach to action quality assessment,” inProceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2019, pp. 304–313. 3

    2019

  25. [25]

    Fineparser: A fine- grained spatio-temporal action parser for human-centric action quality assessment,

    J. Xu, S. Yin, G. Zhao, Z. Wang, and Y . Peng, “Fineparser: A fine- grained spatio-temporal action parser for human-centric action quality assessment,” inProceedings of the IEEE/CVF Conference on computer vision and pattern recognition, 2024, pp. 14 628–14 637. 3, 8, 9

    2024

  26. [26]

    Human-centric fine-grained action qual- ity assessment,

    J. Xu, S. Yin, and Y . Peng, “Human-centric fine-grained action qual- ity assessment,”IEEE Transactions on Pattern Analysis and Machine Intelligence, 2025. 3, 8, 9

    2025

  27. [27]

    Phi: Bridging domain shift in long-term action quality assessment via progressive hierarchical instruction,

    K. Zhou, H. P. Shum, F. W. Li, X. Zhang, and X. Liang, “Phi: Bridging domain shift in long-term action quality assessment via progressive hierarchical instruction,”IEEE Transactions on Image Processing, 2025. 3, 8, 9, 10

    2025

  28. [28]

    Aifit: Automatic 3d human-interpretable feedback models for fitness training,

    M. Fieraru, M. Zanfir, S. C. Pirlea, V . Olaru, and C. Sminchisescu, “Aifit: Automatic 3d human-interpretable feedback models for fitness training,” inProceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2021, pp. 9919–9928. 3

    2021

  29. [29]

    Learning sparse temporal video mapping for action quality assessment in floor gymnastics,

    S. Zahan, G. M. Hassan, and A. Mian, “Learning sparse temporal video mapping for action quality assessment in floor gymnastics,”IEEE Transactions on Instrumentation and Measurement, 2024. 3

    2024

  30. [30]

    Inertial sensor-based analysis of equestrian sports between beginner and professional riders under different horse gaits,

    Z. Wang, J. Li, J. Wang, H. Zhao, S. Qiu, N. Yang, and X. Shi, “Inertial sensor-based analysis of equestrian sports between beginner and professional riders under different horse gaits,”IEEE Transactions on Instrumentation and Measurement, vol. 67, no. 11, pp. 2692–2704,

  31. [31]

    Haptic virtual rehabilitation exercises for poststroke diagnosis,

    A. Alamri, M. Eid, R. Iglesias, S. Shirmohammadi, and A. El Saddik, “Haptic virtual rehabilitation exercises for poststroke diagnosis,”IEEE transactions on instrumentation and measurement, vol. 57, no. 9, pp. 1876–1884, 2008. 3

    2008

  32. [32]

    Sensorized glove for measuring hand finger flexion for rehabilitation purposes,

    M. Borghetti, E. Sardini, and M. Serpelloni, “Sensorized glove for measuring hand finger flexion for rehabilitation purposes,”IEEE Trans- actions on Instrumentation and Measurement, vol. 62, no. 12, pp. 3308– 3314, 2013. 3

    2013

  33. [33]

    Taichi action capture and performance analysis with multi-view rgb cameras,

    J. Li, S. Mo, and Y . Shen, “Taichi action capture and performance analysis with multi-view rgb cameras,”arXiv preprint arXiv:2306.14490,

    work page arXiv

  34. [34]

    Ai-driven tai chi mastery using deep learning framework for movement assessment and personalized training,

    X. Zhao, “Ai-driven tai chi mastery using deep learning framework for movement assessment and personalized training,”Scientific Reports, vol. 15, no. 1, p. 31700, 2025. 3

    2025

  35. [35]

    Development of a motion capture and feedback system for qigong,

    M. Baldinger, K. Lippmann, G. Lisca, and V . Senner, “Development of a motion capture and feedback system for qigong,”Sports Engineering, vol. 28, no. 1, p. 23, 2025. 3

    2025

  36. [36]

    Tui na (or tuina) massage: a minireview of pertinent literature, 1970-2017,

    A. Al-Bedah, G. Ali, T. Abushanab, and N. Qureshi, “Tui na (or tuina) massage: a minireview of pertinent literature, 1970-2017,”Journal of Complementary and Alternative Medical Research, vol. 3, no. 1, pp. 1–14, 2017. 3

    1970

  37. [37]

    Multiview video-based 3-d pose estimation of patients in computer-assisted reha- bilitation environment (caren),

    W. Xu, D. Xiang, G. Wang, R. Liao, M. Shao, and K. Li, “Multiview video-based 3-d pose estimation of patients in computer-assisted reha- bilitation environment (caren),”IEEE Transactions on Human-Machine Systems, vol. 52, no. 2, pp. 196–206, 2022. 3

    2022

  38. [38]

    3d human pose estimation in multi-view operating room videos using differentiable cam- era projections,

    B. G. Gerats, J. M. Wolterink, and I. A. Broeders, “3d human pose estimation in multi-view operating room videos using differentiable cam- era projections,”Computer Methods in Biomechanics and Biomedical Engineering: Imaging & Visualization, vol. 11, no. 4, pp. 1197–1205,

  39. [39]

    Multi- view surgical video action detection via mixed global view attention,

    A. Schmidt, A. Sharghi, H. Haugerud, D. Oh, and O. Mohareri, “Multi- view surgical video action detection via mixed global view attention,” in Medical Image Computing and Computer Assisted Intervention–MICCAI 2021: 24th International Conference, Strasbourg, France, September 27–October 1, 2021, Proceedings, Part IV 24. Springer, 2021, pp. 626–635. 3

    2021

  40. [40]

    A multi-camera, multi-view system for training and skill assessment for robot-assisted surgery,

    A. E. Abdelaal, A. Avinash, M. Kalia, G. D. Hager, and S. E. Salcudean, “A multi-camera, multi-view system for training and skill assessment for robot-assisted surgery,”International journal of computer assisted radiology and surgery, vol. 15, no. 8, pp. 1369–1377, 2020. 3

    2020

  41. [41]

    A data set of human body movements for physical rehabilitation exercises,

    A. Vakanski, H.-p. Jun, D. Paul, and R. Baker, “A data set of human body movements for physical rehabilitation exercises,”Data, vol. 3, no. 1, p. 2, 2018. 3

    2018

  42. [42]

    The kimore dataset: Kinematic assessment of movement and clinical scores for remote monitoring of physical rehabilitation,

    M. Capecci, M. G. Ceravolo, F. Ferracuti, S. Iarlori, A. Monteriu, L. Romeo, and F. Verdini, “The kimore dataset: Kinematic assessment of movement and clinical scores for remote monitoring of physical rehabilitation,”IEEE Transactions on Neural Systems and Rehabilitation Engineering, vol. 27, no. 7, pp. 1436–1448, 2019. 3

    2019

  43. [43]

    Intel- lirehabds (irds)—a dataset of physical rehabilitation movements,

    A. Miron, N. Sadawi, W. Ismail, H. Hussain, and C. Grosan, “Intel- lirehabds (irds)—a dataset of physical rehabilitation movements,”Data, vol. 6, no. 5, p. 46, 2021. 3

    2021

  44. [44]

    A medical low-back pain physical rehabilitation database for human body movement analysis,

    M. Devanne, O. R. Neris, M. Lempereur, A. Thepautet al., “A medical low-back pain physical rehabilitation database for human body movement analysis,” in2024 International Joint Conference on Neural Networks (IJCNN). IEEE, 2024, pp. 1–8. 3

    2024

  45. [45]

    Finerehab: A multi-modality and multi-task dataset for rehabilitation analysis,

    J. Li, J. Xue, R. Cao, X. Du, S. Mo, K. Ran, and Z. Zhang, “Finerehab: A multi-modality and multi-task dataset for rehabilitation analysis,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024, pp. 3184–3193. 3, 4

    2024

  46. [46]

    An exocentric look at egocentric actions and vice versa,

    S. Ardeshir and A. Borji, “An exocentric look at egocentric actions and vice versa,”Computer Vision and Image Understanding, vol. 171, pp. 61–68, 2018. 4

    2018

  47. [47]

    Evaluation specification of clinical prac- tice guidelines of traditional chinese medicine,

    C. A. of Chinese Medicine, “Evaluation specification of clinical prac- tice guidelines of traditional chinese medicine,” China Association of Chinese Medicine, Tech. Rep., January 2021, draft version. 4

    2021

  48. [48]

    The measurement of observer agreement for categorical data,

    J. R. Landis and G. G. Koch, “The measurement of observer agreement for categorical data,”biometrics, pp. 159–174, 1977. 4

    1977

  49. [49]

    Crossvit: Cross-attention multi- scale vision transformer for image classification,

    C.-F. R. Chen, Q. Fan, and R. Panda, “Crossvit: Cross-attention multi- scale vision transformer for image classification,” inProceedings of the IEEE/CVF international conference on computer vision, 2021, pp. 357–

    2021

  50. [50]

    Early vs late fusion in multimodal convolutional neural networks,

    K. Gadzicki, R. Khamsehashari, and C. Zetzsche, “Early vs late fusion in multimodal convolutional neural networks,” in2020 IEEE 23rd international conference on information fusion (FUSION). IEEE, 2020, pp. 1–6. 5 AUTHORet al.: TITLE 15

    2020

  51. [51]

    Deep residual learning for image recognition,

    K. He, X. Zhang, S. Ren, and J. Sun, “Deep residual learning for image recognition,” inProceedings of the IEEE conference on computer vision and pattern recognition, 2016, pp. 770–778. 6, 7

    2016

  52. [52]

    MediaPipe: A Framework for Building Perception Pipelines

    C. Lugaresi, J. Tang, H. Nash, C. McClanahan, E. Uboweja, M. Hays, F. Zhang, C.-L. Chang, M. G. Yong, J. Leeet al., “Mediapipe: A framework for building perception pipelines,”arXiv preprint arXiv:1906.08172, 2019. 6, 7

    work page internal anchor Pith review Pith/arXiv arXiv 1906

  53. [53]

    Zhang, V

    F. Zhang, V . Bazarevsky, A. Vakunov, A. Tkachenka, G. Sung, C.-L. Chang, and M. Grundmann, “Mediapipe hands: On-device real-time hand tracking,”arXiv preprint arXiv:2006.10214, 2020. 6, 7

    work page arXiv 2006

  54. [54]

    Umetrack: Unified multi-view end-to-end hand tracking for vr,

    S. Han, P.-c. Wu, Y . Zhang, B. Liu, L. Zhang, Z. Wang, W. Si, P. Zhang, Y . Cai, T. Hodanet al., “Umetrack: Unified multi-view end-to-end hand tracking for vr,” inSIGGRAPH Asia 2022 conference papers, 2022, pp. 1–9. 6

    2022

  55. [55]

    A compre- hensive survey of action quality assessment: Method and benchmark,

    K. Zhou, R. Cai, L. Wang, H. P. H. Shum, and X. Liang, “A compre- hensive survey of action quality assessment: Method and benchmark,” Pattern Recognition, vol. 179, p. 113933, 2026. 7

    2026

  56. [56]

    Towards a general-purpose foundation model for computational pathology,

    R. J. Chen, T. Ding, M. Y . Lu, D. F. Williamson, G. Jaume, A. H. Song, B. Chen, A. Zhang, D. Shao, M. Shabanet al., “Towards a general-purpose foundation model for computational pathology,”Nature medicine, vol. 30, no. 3, pp. 850–862, 2024. 7

    2024

  57. [57]

    Spatial temporal graph convolutional networks for skeleton-based action recognition,

    S. Yan, Y . Xiong, and D. Lin, “Spatial temporal graph convolutional networks for skeleton-based action recognition,” inProceedings of the AAAI conference on artificial intelligence, vol. 32, no. 1, 2018. 8, 9

    2018

  58. [58]

    High-velocity insertion of acupuncture needle is related to lower level of pain,

    C. S. Yin, J.-H. Kim, and H.-J. Park, “High-velocity insertion of acupuncture needle is related to lower level of pain,”The Journal of Alternative and Complementary Medicine, vol. 17, no. 1, pp. 27–32,

  59. [59]

    A systematic review of skeleton-based action recognition: Methods, chal- lenges, and future directions,

    Y . Liu, R. Liu, Y . Hu, M. Wu, W. Xin, Q. Miao, S. Wu, and L. Li, “A systematic review of skeleton-based action recognition: Methods, chal- lenges, and future directions,”IEEE Transactions on Neural Networks and Learning Systems, 2025. 11

    2025

  60. [60]

    An audiovisual feedback device for compression depth, rate and complete chest recoil can improve the cpr performance of lay persons during self-training on a manikin,

    V . Krasteva, I. Jekova, and J.-P. Didon, “An audiovisual feedback device for compression depth, rate and complete chest recoil can improve the cpr performance of lay persons during self-training on a manikin,” Physiological measurement, vol. 32, no. 6, pp. 687–699, 2011. 12

    2011

  61. [61]

    Hierarchical temporal transformer for 3d hand pose estimation and action recognition from egocentric rgb videos,

    Y . Wen, H. Pan, L. Yang, J. Pan, T. Komura, and W. Wang, “Hierarchical temporal transformer for 3d hand pose estimation and action recognition from egocentric rgb videos,” inProceedings of the IEEE/CVF Confer- ence on Computer Vision and Pattern Recognition, 2023, pp. 21 243– 21 253. 13

    2023

  62. [62]

    Geometric features informed multi-person human-object interaction recognition in videos,

    T. Qiao, Q. Men, F. W. Li, Y . Kubotani, S. Morishima, and H. P. Shum, “Geometric features informed multi-person human-object interaction recognition in videos,” inEuropean Conference on Computer Vision. Springer, 2022, pp. 474–491. 13

    2022

  63. [63]

    Geometric fea- tures enhanced human-object interaction detection,

    M. Zhu, E. S. Ho, S. Chen, L. Yang, and H. P. Shum, “Geometric fea- tures enhanced human-object interaction detection,”IEEE Transactions on Instrumentation and Measurement, 2024. 13

    2024

  64. [64]

    Adaptive graph learning from spatial information for surgical workflow anticipation,

    F. X. Zhang, J. Deng, R. Lieck, and H. P. Shum, “Adaptive graph learning from spatial information for surgical workflow anticipation,” IEEE Transactions on Medical Robotics and Bionics, 2024. 13

    2024