Pith. sign in

REVIEW 4 major objections 6 minor 50 references

Real-Time Feedback and Benchmark Dataset for Isometric Pose Evaluation

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

Pith's one-line read The paper claims that a simple angle-histogram classifier matches or beats a two-stream graph convolutional network on isometric pose classification while making far fewer low-confidence mistakes, and that a new three-part metric is…

desk verdict A genuinely new isometric-exercise dataset with a sensible confidence metric, but the safety claim rests on unvalidated labels and a withheld release. read the letter →

arxiv 2506.11774 v1 pith:EWBAC4A4 submitted 2025-06-13 cs.CV cs.AIcs.HC

classification cs.CVcs.AIcs.HC
keywords isometricexerciseposeevaluationdatasetskeleton-basedclassificationgraphconvolutionalnetworkmodelconfidencereal-timefeedbackhumanestimation
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

Isometric exercises are done at home without a trainer, so a system that watches a user and says how to fix a pose could prevent injury and dropout. The paper's central proposal is that this feedback can be built on a dataset that includes both correct executions and common mistakes, rather than correct poses only. It introduces IMCD, over 3,600 video clips of six isometric poses performed by about ten male subjects, each pose with one correct and two incorrect variations. Benchmarking three model families on this dataset, it finds that a simple classifier built on hold-phase joint-angle histograms matches or beats a two-stream graph convolutional network on accuracy while being far more confident about its mistakes, measured by a new three-part metric. The paper argues that this confidence difference, not raw accuracy, is the property that matters when the feedback is delivered to a real exerciser.

What carries the argument

The load-bearing object is the hold-phase angle histogram. For each rep, the method computes joint angles $\theta_{ijk}$ between joint triplets, builds a histogram of each angle over the rep's frames, and takes the highest-count bin center as the feature; these features feed an MLP. Rep discrimination is done with peak prominence on a joint's motion signal, retaining peaks above a prominence threshold $\tau = 0.2$. The other mechanism is the three-part metric, which separates predictions into confident correct/incorrect decisions (M1), confident mistake classifications (M2), and low-confidence mistakes (M3), with M3 counted as the fraction of incorrect examples where the model's top incorrect-class probability stays below 0.5. That metric carries the paper's central safety argument: a model can have high multiclass accuracy yet be unsafe because a large share of its mistakes are low-confidence confusions.

What would settle it

Record a new set of exercisers performing the same six isometric poses without being told what mistake to make, have certified trainers label the errors, and score the angle-based classifier and 2s-AGCN with the three-part metric; if the angle-based classifier's low-confidence mistake fraction (M3) rises to or above the graph network's, the paper's safety advantage for the simple model is refuted.

Watch

Extended reading notes

Core claim

The paper claims that pose correctness in isometric exercises can be judged from the distribution of joint angles during the static hold phase, and that a multi-layer perceptron trained on those angle histograms is the most reliable of the tested options for real-time feedback. On the released IMCD, this angle-based classifier reaches weighted F1 scores between 0.796 and 0.952 across the six poses. Against 2s-AGCN, which scores a higher multiclass F1 on the demonstration Tree Pose (0.956 vs 0.947), the angle-based classifier has a better binary correct-versus-incorrect F1 (0.959 vs 0.870) and a much smaller fraction of low-confidence mistaken predictions (M3 = 0.21% vs 12.9%). The discovery is therefore twofold: a simple geometric feature is competitive with learned graph representations for isometric poses, and the proposed three-part metric exposes a safety-relevant gap that multiclass F1 hides.

Load-bearing premise

The central safety conclusion assumes the mistakes recorded in IMCD match the mistakes real exercisers actually make, because each exerciser was simply told which mistake to simulate and no expert validation or inter-annotator agreement is reported.

Editorial extensions

If this is right

  • A dataset pairing correct and incorrect isometric variations makes it possible to train mistake-aware classifiers instead of relying on manually set angular thresholds.
  • Angle-histogram features extracted from the hold phase alone are sufficient for competitive classification of six isometric poses, supporting near real-time feedback.
  • Multiclass F1 should not be the sole yardstick in exercise feedback; the three-part metric reveals that a graph network's superior multiclass score coexists with far higher uncertainty about wrong predictions.
  • The three-part metric can be tuned with $F_\beta$ to weight precision, which is the safety-relevant direction when avoiding classifying an incorrect rep as correct.
  • Graph convolutional models occasionally edge out the angle-based classifier on specific poses, so hybrid approaches are a reasonable next step rather than a settled conclusion.

Reading between the lines

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

  • Beyond the paper: the safety gap reported for 2s-AGCN could shrink or vanish on a dataset of naturally occurring (rather than instructed) mistakes, so the conclusion that simpler models are safer should be re-tested before deployment.
  • Beyond the paper: the three-part metric is domain-agnostic and could be applied to any classifier where low-confidence errors are costly, such as medical screening or industrial inspection, though the authors do not claim this.
  • Beyond the paper: because IMCD was recorded with only male participants and six poses, an immediate testable extension is whether the angle-based classifier's edge persists across genders, ages, and exercise families; the paper acknowledges this limitation.
  • Beyond the paper: the angle-histogram feature discards temporal ordering entirely, so adding a simple temporal prior (e.g., per-rep phase or smoothing across frames) might close the gap to graph models while preserving low uncertainty, which the paper does not investigate.
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 introduces IMCD, a video dataset of six isometric exercises with one correct and two incorrect classes per exercise, collected from roughly ten male participants using mobile cameras. It proposes an angle-histogram feature plus MLP classifier, reuses an Interaction-Network motion-prediction baseline from the authors' prior work, and compares against 2s-AGCN. It also defines a three-part metric (M1/M2/M3) intended to separate correct/incorrect discrimination, confident mistake classification, and low-confidence mistakes. The central empirical claims are that the angle-based classifier performs comparably to or better than 2s-AGCN on weighted F1 (Table 3) and exhibits far lower uncertainty (M3=0.21% vs 12.9%) on Tree Pose (Table 4), making it safer for real-time feedback. The dataset and code are promised for release post-acceptance.

Significance. If the labels of IMCD can be trusted, the dataset would be a useful community resource because existing isometric datasets mostly contain correct poses only. The angle-based baseline is a simple, cheaply deployable alternative, and the M1/M2/M3 decomposition addresses a real gap in evaluation for feedback systems. The paper also benefits from testing a graph-based model on a domain where it is not the default. However, the strength of these contributions is currently conditional: the ground-truth labels are unvalidated, the benchmark protocol is under-specified, and the headline safety result rests on a single pose. With additional validation and reproducible reporting, this could become a solid dataset/benchmark contribution.

major comments (4)
  1. [§3 (Dataset curation)] The ground-truth labels are the load-bearing component of every result in the paper, yet Section 3 reports only that exercisers 'were informed about the specific mistake to demonstrate' and 'had the freedom to simulate these mistakes within the defined parameters of each mistake category.' No expert validation, inter-annotator agreement, or comparison with naturally occurring mistakes is reported, and Section 7 does not flag this gap. If labels are noisy or if instructed mistakes are not representative of real user errors, every F1, M1, M2, and M3 value in Tables 3 and 4 changes, and the safety conclusion in Section 6 loses its basis. Please add a label-validation protocol (e.g., expert annotation of a held-out subset, agreement statistics, per-class error analysis) and report label statistics per subject and per class.
  2. [§5, Table 3] Table 3 reports a single weighted F1 per pose per method, with no standard deviation, no train/validation/test split, no class-wise F1, and no significance test. In the absence of a described protocol, the reader cannot know whether the differences between methods (e.g., 0.952 vs 0.914 for Cobra, 0.802 vs 0.830 for Superman) are meaningful. Please specify the evaluation protocol (cross-validation folds or fixed split, number of runs, data augmentation if any) and report mean ± std and per-class results, or at minimum provide the confusion matrices.
  3. [§4.1–§4.3, §5] The compared methods are not fully specified. For the angle-based classifier, n (histogram bins), the set of joint triplets, the MLP architecture and training details, and the precise role of the 'one standard deviation' range in Section 4.1 are not given. For 2s-AGCN, training epochs, batch size, optimizer, and input normalization are absent. The motion-prediction baseline is deferred entirely to reference [20]. Since the dataset and code will only be released post-acceptance, the benchmark is not reproducible in its present form. Please add an appendix with full hyperparameters and, ideally, release the dataset/code to reviewers at revision time.
  4. [§6, Table 4, Algorithm 1] The three-part metric is demonstrated only on Tree Pose, yet the text generalizes: 'the angle-based approach achieves the best binary F1 score and demonstrates exceptional reliability' (Section 6). As a result, the central safety claim is not supported for the other five poses. Also, M3 in Algorithm 1 is computed as the fraction of incorrect examples whose argmax is an incorrect class with max probability below τ; examples where the model predicts 'correct' are excluded from the numerator but included in the denominator. This makes M3 a conditional low-confidence rate rather than a general uncertainty measure, and its values are threshold-dependent. Please report M1–M3 for all poses and clarify the interpretation of M3, including sensitivity to τ.
minor comments (6)
  1. [Abstract, §1, §2] There are typos in the text: 'equipments' should be 'equipment', and Section 2 contains 'poset-exercise' and 'ranger' instead of the intended words; these should be corrected.
  2. [Table 1] The IMCD row in Table 1 marks 'Available' as 'Yes', while Section 3 says the full dataset will be released 'post-acceptance'; please reconcile these statements and clarify what is currently accessible.
  3. [§4.1, Eq. (4)] The notation 'argmax_{B_k} H_{\theta_i}(b_k)' is ambiguous because two different variables are both indexed by k; please clarify that the feature is the bin center associated with the maximum count.
  4. [Algorithm 1] The symbol 'maxP(incorrect)_i' is not defined; please define it as the maximum probability over the incorrect classes for example i.
  5. [§4.1] The relationship between the 'one standard deviation' correct-angle range and the trained MLP classifier is unclear; state whether the range is used for feature normalization, label assignment, or evaluation.
  6. [Table 4] Since only Tree Pose is shown, either add a supplementary table for all poses or explicitly label the result as a case study.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the benchmark is a dataset evaluation, the sole self-citation is a non-load-bearing baseline, and no result reduces by construction to its inputs.

full rationale

The paper's central claims are the release of the IMCD dataset, the benchmarking of three classifiers on it, and the proposal of a three-part evaluation metric. Dataset-and-benchmark papers are not circular merely because the benchmark is computed on the authors' own data; the models are trained and evaluated on the same released data under a standard classification setup, and the results in Tables 3 and 4 are empirical rather than definitional. The only self-citation is the motion-prediction baseline of Jaiswal et al. [20], detailed in Section 4.2 and referenced as the source of the Interaction-Network plus Fourier-deviation pipeline. That baseline is one of three compared methods, is not used to define the dataset or the metric, and performs worst in Table 3, so none of the paper's conclusions depend on it. The proposed three-part metric (Algorithm 1) is a definition applied to model outputs; it does not preordain that the angle-based classifier wins the M3 comparison, and the lower M3 value for the angle-based model is an empirical outcome, not a fitted input. Concerns about the lack of expert validation of the simulated mistake labels affect ground-truth quality and generalizability, and are acknowledged only partially in Section 7; this is a validity risk, not circular reasoning. No equation or claim in the paper reduces by construction to another, so the appropriate score is 0.

Assumptions & free parameters 5 free parameters · 4 assumptions · 0 invented entities

The central empirical claims depend on the quality of human-assigned labels, the reliability of undeclared keypoint extraction, and the transfer of the authors' earlier motion-prediction pipeline. Numbers chosen by hand (prominence threshold 0.2, confidence threshold 0.5, histogram bin count, MLP architecture) enter the pipeline without sensitivity analysis. No new physical entities are introduced.

free parameters (5)
  • confidence threshold τ in Algorithm 1 = 0.5
    Splits confident versus uncertain incorrect predictions in M2 and M3; the paper's conclusion that the angle-based method is safer depends on this threshold and no sensitivity analysis is given.
  • prominence threshold τ in Eq. 1 = 0.2
    Used to detect exercise repetitions from the joint motion signal; chosen without sensitivity analysis or manual validation.
  • histogram bin count n for angle features
    Equation 3 defines a histogram over n bins but n is never reported, so the angle feature vector is underspecified.
  • correct angle range (±1 standard deviation)
    Section 4.1 uses mean ± 1 SD as the standard correct range; although the final classifier is an MLP, this range is used for difficulty framing and conceptual grounding.
  • MLP hyperparameters
    The angle-based classifier is an MLP, but architecture, hidden units, training epochs, and regularization are not reported.
assumptions (4)
  • domain assumption Exercise labels are reliable and representative of common real-world isometric mistakes.
    Invoked throughout Section 3 and used as ground truth for all tables; no inter-annotator agreement, expert validation protocol, or label error analysis is reported.
  • domain assumption The pose keypoint extraction used by all methods is accurate and consistent across subjects, distances, and indoor/outdoor settings.
    Section 4 states a skeleton keypoint representation but never names the keypoint detector or its error characteristics; the angle and graph methods depend on it.
  • domain assumption Repetition boundaries found by local-minimum valleys around prominence peaks correctly segment each exercise rep for all six poses.
    Section 4, Equation 1 uses a prominence threshold without validation against manual segmentation or ground-truth rep boundaries.
  • ad hoc to paper The authors' previous Interaction-Network pipeline transfers to isometric exercises without modification.
    Section 4.2 defers all details to Jaiswal et al. [20], a same-author paper, and uses it as a baseline; the transfer is asserted, not validated here.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Real-Time Feedback and Benchmark Dataset for Isometric Pose Evaluation." pith.science (2026). https://pith.science/paper/EWBAC4A4

@misc{pith2026250611774,
  author       = {Pith},
  title        = {Pith review of: Real-Time Feedback and Benchmark Dataset for Isometric Pose Evaluation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/EWBAC4A4}},
  note         = {Machine review of arXiv:2506.11774}
}
read the original abstract

Isometric exercises appeal to individuals seeking convenience, privacy, and minimal dependence on equipments. However, such fitness training is often overdependent on unreliable digital media content instead of expert supervision, introducing serious risks, including incorrect posture, injury, and disengagement due to lack of corrective feedback. To address these challenges, we present a real-time feedback system for assessing isometric poses. Our contributions include the release of the largest multiclass isometric exercise video dataset to date, comprising over 3,600 clips across six poses with correct and incorrect variations. To support robust evaluation, we benchmark state-of-the-art models-including graph-based networks-on this dataset and introduce a novel three-part metric that captures classification accuracy, mistake localization, and model confidence. Our results enhance the feasibility of intelligent and personalized exercise training systems for home workouts. This expert-level diagnosis, delivered directly to the users, also expands the potential applications of these systems to rehabilitation, physiotherapy, and various other fitness disciplines that involve physical motion.

Figures

Figures reproduced from arXiv: 2506.11774 by the authors.

Figure 1
Figure 1. A sample from IMCD: Video Motion Trail for Tree Pose for three classes [PITH_FULL_IMAGE:figures/full_fig_p005_1.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

50 extracted references · 48 canonical work pages

  1. [20]

    Using learnable physics for real-time exercise form recommendations

    Abhishek Jaiswal, Gautam Chauhan, and Nisheeth Srivastava. Using learnable physics for real-time exercise form recommendations. In Proceedings of the 17th ACM Con- ference on Recommender Systems , RecSys ’23, page 688–695, New York, NY , USA,

  2. [1]

    https://en

    Topographic prominence - Wikipedia — en.wikipedia.org. https://en. wikipedia.org/wiki/Topographic_prominence. [Accessed 15-05- 2025]

  3. [2]

    Yoga pose estimation and feedback generation using deep learning

    Vivek Anand Thoutam, Anugrah Srivastava, Tapas Badal, Vipul Kumar Mishra, GR Sinha, Aditi Sakalle, Harshit Bhardwaj, and Manish Raj. Yoga pose estimation and feedback generation using deep learning. Computational Intelligence and Neuro- science, 2022(1):4311350, 2022. AUTHORS:: REAL-TIME FEEDBACK FOR ISOMETRIC POSE EV ALUA TION 11

  4. [3]

    Validation of Human Pose Estimation and Human Mesh Recovery for Extracting Clinically Relevant Motion Data from Videos

    Kai Armstrong, Alexander Rodrigues, Alexander P Willmott, Lei Zhang, and Xujiong Ye. Validation of human pose estimation and human mesh recovery for extracting clinically relevant motion data from videos. arXiv preprint arXiv:2503.14760, 2025

  5. [4]

    An evidence-based guide to the efficacy and safety of isometric resis- tance training in hypertension and clinical implications

    Biggie Baffour-Awuah, Melissa J Pearson, Gudrun Dieberg, Jonathan D Wiles, and Neil A Smart. An evidence-based guide to the efficacy and safety of isometric resis- tance training in hypertension and clinical implications. Clinical Hypertension, 29(1): 9, 2023

  6. [5]

    Inter- action networks for learning about objects, relations and physics

    Peter Battaglia, Razvan Pascanu, Matthew Lai, Danilo Jimenez Rezende, et al. Inter- action networks for learning about objects, relations and physics. Advances in neural information processing systems, 29, 2016

  7. [6]

    Recommended exercises for lower back pain on youtube videos; are they useful or harmful? Health Policy and Technol- ogy, 13(3):100891, 2024

    Rabia Aydogan Baykara and Nevsun Pıhtılı Ta¸ s. Recommended exercises for lower back pain on youtube videos; are they useful or harmful? Health Policy and Technol- ogy, 13(3):100891, 2024

  8. [7]

    Yog-guru: Real-time yoga pose correction system using deep learning methods

    Ajay Chaudhari, Omkar Dalvi, Onkar Ramade, and Dayanand Ambawade. Yog-guru: Real-time yoga pose correction system using deep learning methods. In 2021 interna- tional conference on communication information and computing technology (ICCICT), pages 1–6. IEEE, 2021

Show all 50 references
  1. [8]

    Yoga posture recognition for self-training

    Hua-Tsung Chen, Yu-Zhen He, Chun-Chieh Hsu, Chien-Li Chou, Suh-Yin Lee, and Bao-Shuh P Lin. Yoga posture recognition for self-training. In MultiMedia Modeling: 20th Anniversary International Conference, MMM 2014, Dublin, Ireland, January 6- 10, 2014, Proceedings, Part I 20, pa...

  2. [9]

    Computer-assisted yoga training system

    Hua-Tsung Chen, Yu-Zhen He, and Chun-Chieh Hsu. Computer-assisted yoga training system. Multimedia Tools and Applications, 77:23969–23991, 2018

  3. [10]

    Pose trainer: correcting exercise posture using pose estimation

    Steven Chen and Richard R Yang. Pose trainer: correcting exercise posture using pose estimation. arXiv preprint arXiv:2006.11718, 2020

  4. [11]

    Trends over 5 decades in us occupation-related physical activity and their associations with obesity

    Timothy S Church, Diana M Thomas, Catrine Tudor-Locke, Peter T Katzmarzyk, Conrad P Earnest, Ruben Q Rodarte, Corby K Martin, Steven N Blair, and Claude Bouchard. Trends over 5 decades in us occupation-related physical activity and their associations with obesity. PloS one, 6(...

  5. [12]

    Pose tutor: an explainable system for pose correction in the wild

    Bhat Dittakavi, Divyagna Bavikadi, Sai Vikas Desai, Soumi Chakraborty, Nishant Reddy, Vineeth N Balasubramanian, Bharathi Callepalli, and Ayon Sharma. Pose tutor: an explainable system for pose correction in the wild. InProceedings of the IEEE/CVF conference on computer vision...

  6. [13]

    Pyskl: Towards good prac- tices for skeleton action recognition

    Haodong Duan, Jiaqi Wang, Kai Chen, and Dahua Lin. Pyskl: Towards good prac- tices for skeleton action recognition. In Proceedings of the 30th ACM International Conference on Multimedia, pages 7351–7354, 2022

  7. [14]

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

    Mihai Fieraru, Mihai Zanfir, Silviu Cristian Pirlea, Vlad Olaru, and Cristian Sminchis- escu. Aifit: Automatic 3d human-interpretable feedback models for fitness training. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9919–9928, 20...

  8. [15]

    Short: Deep learn- ing approach to skeletal performance evaluation of physical therapy exercises

    Bhanu Garg, Alexander Postlmayr, Pamela Cosman, and Sujit Dey. Short: Deep learn- ing approach to skeletal performance evaluation of physical therapy exercises. In Pro- ceedings of the 8th ACM/IEEE International Conference on Connected Health: Appli- cations, Systems and Engin...

  9. [16]

    Performance monitoring for exercise movements using mobile cameras

    Morteza Mohammady Gharasuie, Noah Jennings, and Shubham Jain. Performance monitoring for exercise movements using mobile cameras. InProceedings of the Work- shop on Body-Centric Computing Systems, pages 1–6, 2021

  10. [17]

    Practice or per- formance? a content analysis of yoga-related videos on instagram

    Angela Hinz, Kate E Mulgrew, Tamara De Regt, and Geoff Lovell. Practice or per- formance? a content analysis of yoga-related videos on instagram. Body image, 39: 175–183, 2021

  11. [18]

    Three- dimensional cnn-inspired deep learning architecture for yoga pose recognition in the real-world environment

    Shrajal Jain, Aditya Rustagi, Sumeet Saurav, Ravi Saini, and Sanjay Singh. Three- dimensional cnn-inspired deep learning architecture for yoga pose recognition in the real-world environment. Neural Computing and Applications, 33:6427–6441, 2021

  12. [19]

    Benchmarking reliability of deep learn- ing models for pathological gait classification

    Abhishek Jaiswal and Nisheeth Srivastava. Benchmarking reliability of deep learn- ing models for pathological gait classification. In Machine Learning for Healthcare Conference. PMLR, 2024

  13. [21]

    Virtual exercise trends: A guide to the growing popularity of online exercise

    Northwest Indiana Journal. Virtual exercise trends: A guide to the growing popularity of online exercise. https://nwijournal.com/ virtual-exercise-trends-a-guide-to-the-growing-popularity-of-online-exercise/ ,

  14. [22]

    3dyoga90: A hierarchical video dataset for yoga pose understanding

    Seonok Kim. 3dyoga90: A hierarchical video dataset for yoga pose understanding. arXiv preprint arXiv:2310.10131, 2023

  15. [23]

    Yoga pose classification using deep learning

    Shruti Kothari. Yoga pose classification using deep learning. 2020

  16. [24]

    When adults don’t exercise: Behavioral strategies to increase physical activity in sedentary middle-aged and older adults

    Margie E Lachman, Lewis Lipsitz, James Lubben, Carmen Castaneda-Sceppa, and Alan M Jette. When adults don’t exercise: Behavioral strategies to increase physical activity in sedentary middle-aged and older adults. Innovation in aging, 2(1):igy007, 2018

  17. [25]

    When you don’t get what you want—and it’s really hard: Exploring motivational contributions to exercise dropout

    Heather K Larson, Kimberley Mcfadden, Tara-Leigh F McHugh, Tanya R Berry, and Wendy M Rodgers. When you don’t get what you want—and it’s really hard: Exploring motivational contributions to exercise dropout. Psychology of Sport and Exercise, 37: 59–66, 2018

  18. [26]

    A posture evaluation system for fitness videos based on recurrent neural network

    An-Lun Liu and Wei-Ta Chu. A posture evaluation system for fitness videos based on recurrent neural network. In 2020 International Symposium on Computer, Consumer and Control (IS3C), pages 185–188. IEEE, 2020. AUTHORS:: REAL-TIME FEEDBACK FOR ISOMETRIC POSE EV ALUA TION 13

  19. [27]

    Dis- entangling and unifying graph convolutions for skeleton-based action recognition

    Ziyu Liu, Hongwen Zhang, Zhenghao Chen, Zhiyong Wang, and Wanli Ouyang. Dis- entangling and unifying graph convolutions for skeleton-based action recognition. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 143–152, 2020

  20. [28]

    3DYogaSeg: A New Dataset and Benchmark for Skeleton-Based Action Recognition and Segmentation in yoga videos

    Edoardo Marchetti. 3DYogaSeg: A New Dataset and Benchmark for Skeleton-Based Action Recognition and Segmentation in yoga videos. PhD thesis, Politecnico di Torino, 2024

  21. [29]

    Physical exer- cise form correction using neural networks

    Cristian Militaru, Maria-Denisa Militaru, and Kuderna-Iulian Benta. Physical exer- cise form correction using neural networks. In Companion Publication of the 2020 International Conference on Multimodal Interaction, pages 240–244, 2020

  22. [30]

    Classification of yoga asanas from a single image by learning the 3d view of human poses

    Chirumamilla Nagalakshmi and Snehasis Mukherjee. Classification of yoga asanas from a single image by learning the 3d view of human poses. Digital techniques for heritage presentation and preservation, pages 37–49, 2021

  23. [31]

    Spatio-temporal pyramid graph convolutions for human action recognition and postural assessment

    Behnoosh Parsa, Behzad Dariush, et al. Spatio-temporal pyramid graph convolutions for human action recognition and postural assessment. InProceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 1080–1090, 2020

  24. [32]

    Erica: enabling real-time mistake detection & corrective feedback for free-weights exercises

    Meera Radhakrishnan, Darshana Rathnayake, Ong Koon Han, Inseok Hwang, and Archan Misra. Erica: enabling real-time mistake detection & corrective feedback for free-weights exercises. InProceedings of the 18th Conference on Embedded Networked Sensor Systems, pages 558–571, 2020

  25. [33]

    Review of the quality of youtube videos recommending exercises for the covid-19 lockdown

    Alvaro Manuel Rodriguez-Rodriguez, Maria Blanco-Diaz, Marta de la Fuente-Costa, Sergio Hernandez-Sanchez, Isabel Escobio-Prieto, and Jose Casaña. Review of the quality of youtube videos recommending exercises for the covid-19 lockdown. Inter- national journal of environmental ...

  26. [34]

    Effects of correctness and suggestive feedback on learning with an autonomous virtual trainer

    Xiumin Shang, Marcelo Kallmann, and Ahmed Sabbir Arif. Effects of correctness and suggestive feedback on learning with an autonomous virtual trainer. In Companion Proceedings of the 24th International Conference on Intelligent User Interfaces, pages 93–94, 2019

  27. [35]

    Two-stream adaptive graph convolutional networks for skeleton-based action recognition

    Lei Shi, Yifan Zhang, Jian Cheng, and Hanqing Lu. Two-stream adaptive graph convolutional networks for skeleton-based action recognition. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages 12026– 12035, 2019

  28. [36]

    Health and functional benefits of yoga practice in times of the covid-19 coronavirus pandemic

    Olga Ilona Smole ´nska and Marta Pilarska. Health and functional benefits of yoga practice in times of the covid-19 coronavirus pandemic. Pedagogy and Psychology of Sport, 7(2):21–38, 2021

  29. [37]

    Human action recognition from various data modalities: A review

    Zehua Sun, Qiuhong Ke, Hossein Rahmani, Mohammed Bennamoun, Gang Wang, and Jun Liu. Human action recognition from various data modalities: A review. IEEE transactions on pattern analysis and machine intelligence, 45(3):3200–3225, 2022

  30. [38]

    Bodylights: Open-ended augmented feedback to support training towards a correct exercise execution

    Laia Turmo Vidal, Hui Zhu, and Abraham Riego-Delgado. Bodylights: Open-ended augmented feedback to support training towards a correct exercise execution. In Pro- ceedings of the 2020 CHI Conference on Human Factors in Computing Systems, pages 1–14, 2020. 14 AUTHORS:: REAL-TIME...

  31. [39]

    Yoga-82: a new dataset for fine-grained classification of human poses

    Manisha Verma, Sudhakar Kumawat, Yuta Nakashima, and Shanmuganathan Raman. Yoga-82: a new dataset for fine-grained classification of human poses. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition workshops , pages 1038–1039, 2020

  32. [40]

    Ai coach: Deep human pose estimation and analysis for personalized athletic training assistance

    Jianbo Wang, Kai Qiu, Houwen Peng, Jianlong Fu, and Jianke Zhu. Ai coach: Deep human pose estimation and analysis for personalized athletic training assistance. In Proceedings of the 27th ACM International Conference on Multimedia , pages 374– 382, 2019

  33. [41]

    A novel method for the performance of isometric exercise in the home

    Jonathan D Wiles, Damian Coleman, Martin Dunford, and Ian Swaine. A novel method for the performance of isometric exercise in the home. Journal of sports sciences, 23 (8):795–803, 2005

  34. [42]

    Home-based isometric ex- ercise training induced reductions resting blood pressure

    Jonathan D Wiles, Natalie Goldring, and Damian Coleman. Home-based isometric ex- ercise training induced reductions resting blood pressure. European journal of applied physiology, 117:83–93, 2017

  35. [43]

    Nearly 1.8 billion adults at risk of disease from not doing enough physical activity

    World Health Organization. Nearly 1.8 billion adults at risk of disease from not doing enough physical activity. https: //www.who.int/news/item/26-06-2024-nearly-1. 8-billion-adults-at-risk-of-disease-from-not-doing-enough-physical-activity ,

  36. [44]

    A computer vision-based yoga pose grading approach using con- trastive skeleton feature representations

    Yubin Wu, Qianqian Lin, Mingrun Yang, Jing Liu, Jing Tian, Dev Kapil, and Laura Vanderbloemen. A computer vision-based yoga pose grading approach using con- trastive skeleton feature representations. In Healthcare, volume 10, page 36. MDPI, 2021

  37. [45]

    Real-time yoga recognition using deep learning

    Santosh Kumar Yadav, Amitojdeep Singh, Abhishek Gupta, and Jagdish Lal Raheja. Real-time yoga recognition using deep learning. Neural computing and applications, 31:9349–9361, 2019

  38. [46]

    Accessed: 2024-12-20

  39. [47]

    Yoga survey: Statistics and insights

    DoYou Yoga. Yoga survey: Statistics and insights. https://www. doyou.com/yoga-survey/#:~:text=Key%20Finding%20%238: %20More%20than,Others%20(14.6%25), 2024. Accessed: 2024-12-20

  40. [48]

    3d pose-based feedback for physical exercises

    Ziyi Zhao, Sena Kiciroglu, Vinzant, Hugues, Cheng, Yuan, Katircioglu, Isinsu, Salz- mann, Mathieu, and Fua, Pascal. 3d pose-based feedback for physical exercises. In Proceedings of the Asian Conference on Computer Vision, pages 1316–1332, 2022

  41. [49]

    Spatial temporal graph convolutional net- works for skeleton-based action recognition

    Sijie Yan, Yuanjun Xiong, and Dahua Lin. Spatial temporal graph convolutional net- works for skeleton-based action recognition. In Proceedings of the AAAI conference on artificial intelligence, volume 32, 2018

  42. [2023]

    ISBN 9798400702419

    Association for Computing Machinery. ISBN 9798400702419. doi: 10.1145/ 3604915.3608816. URL https://doi.org/10.1145/3604915.3608816

Pith tools

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