Pith. sign in

REVIEW 4 major objections 5 minor 56 references

(LiFT) Lightweight Fitness Transformer: A language-vision model for Remote Monitoring of Physical Training

T0 review · 4 major / 5 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 A single vision-language transformer takes 3D skeletal keypoints from RGB video and reports both the exercise name and the repetition count across more than 1,900 movements, scoring 76.5% on detection and 85.3% off-by-one on rep counting.

desk verdict A useful dataset and a sensible system, but the headline 76.5% exercise-detection 'accuracy' is recall, not accuracy, and the rep-counting numbers lack trivial baselines. read the letter →

arxiv 2506.06480 v1 pith:HQMHQWF4 submitted 2025-06-06 cs.CV

A. Postlmayr , P. Cosman , S. Dey This is my paper
classification cs.CV
keywords exercisedetectionrepetitioncountingvision-languagetransformerskeleton-basedactionrecognition3DhumanposeestimationfitnessmonitoringmotionimageRGBvideo
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

LiFT sets out to make automated fitness tracking private, cheap, and scalable by replacing raw video with skeletal keypoints. It is a vision-and-language transformer that takes a motion image built from 3D poses and outputs both the exercise name and the repetition count, treating the exercise label as a bag of weighted words plus an integer. To train it, the authors assembled Olympia, a dataset of 7,618 videos covering more than 1,900 exercises from existing datasets and open-source footage. On its test split, LiFT reports 76.5% exercise-detection and 85.3% off-by-one repetition-counting scores using only RGB video, which the paper presents as the first skeleton-based vision-language model to handle both tasks at this scale. If correct, this would support at-home fitness coaching on a smartphone without sending video to the cloud.

What carries the argument

The load-bearing mechanism is the 'motion image': a smoothed skeleton sequence of 17 keypoints, split into five kinematic chains (spine and head, left leg, right leg, left arm, right arm), each interpolated to 64 evenly spaced points, then concatenated into a single image-like tensor. This tensor, together with a task question such as 'Which exercise was performed?' or 'How many repetitions were done?', is fed into a ViLT, a vision-and-language transformer, whose classification head predicts a multi-hot vector over a shared vocabulary of 441 exercise-description words plus the integers 1 to 30. Word categories are weighted from 0 to 1.0 so core action words and body parts dominate while equipment words (which 3D pose estimation cannot see) are down-weighted, making it possible for one model to answer both questions: the exercise label is read off as the set of confident predicted words, and the repetition count as the single most confident integer prediction.

What would settle it

Rescore the model on the same Olympia test split using top-1 exact label match or precision (matched words divided by predicted words); if exact-match accuracy is substantially below 76.5%, the headline accuracy claim does not hold.

Watch

Extended reading notes

Core claim

The paper's central claim is that a single lightweight vision-language transformer can jointly recognize exercises and count repetitions from skeletal motion alone, at a scale far beyond prior fitness-analysis models, which typically handle fewer than 30 exercises and rely on handcrafted templates. The evidence is LiFT's performance on Olympia: 76.5% exercise detection using a partial-credit word-overlap score, and 85.3% off-by-one repetition-counting accuracy with a mean absolute error of 0.645, all from RGB video converted to 17 keypoints by a 3D human-pose-estimation model. The paper argues that word-level multi-class prediction, rather than global image-text alignment, is what lets the model capture fine-grained sub-motion-to-word relationships, and that the same model generalizes to unseen exercises through fine-tuning and few-shot adaptation.

Load-bearing premise

The public claim of 76.5% exercise-detection accuracy rests on a score that counts only matched ground-truth words and never penalizes extra or wrong predictions, so the number does not mean what 'accuracy' usually means.

Editorial extensions

If this is right

  • The 433 MB model-weight footprint is about 60 times smaller than the 25.6 GB needed by the large motion-language LLMs the paper compares against, which the authors cite as evidence the architecture is suited for on-device fitness tracking.
  • Fine-tuning on Fit3D lifts exercise detection to 89% and off-by-one rep counting to 99.6%, showing the model adapts to a new exercise set with modest amounts of labeled video.
  • The 91.27% accuracy on the idiosyncratic 'warmupX' labels indicates the model learns global motion properties alongside word-level sub-motions, which helps explain why it can handle exercises without standard names.
  • Substituting Mediapipe for MotionBERT as the pose estimator lowers exercise detection by 14.5 percentage points and rep-count OBO by 15.9 points, setting a target for what mobile-grade pose estimation must deliver.

Reading between the lines

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

  • Because the reported exercise-detection score is recall with no penalty for extra words, the headline 76.5% is an upper bound on any precision-style or exact-match accuracy, and a deployment would need exact-match or top-k scoring to be meaningful.
  • The word-level vocabulary opens a path to open-set recognition: exercises whose names are novel combinations of known words, such as 'weighted reverse lunge', could be recognized without retraining, going beyond the 1,900 curated labels.
  • The deliberately low weight on equipment words (0.1) means the model cannot tell apart exercises that differ only by equipment; adding a lightweight object detector could recover that information and likely lift detection on equipment-dependent labels.
  • The modest zero-shot scores on rehabilitation datasets (16.7% to 20.8% detection) suggest the fitness-trained vocabulary does not transfer cleanly to clinical movements, so a rehabilitation-specific vocabulary or additional pose data would be needed before clinical deployment.
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 / 5 minor

Summary. The paper presents LiFT, a vision-language transformer that consumes 3D skeletal keypoints extracted from RGB video by MotionBERT and recast as motion images, and is trained on Olympia, a newly assembled dataset of 7,618 videos, to jointly perform exercise detection (multi-label word prediction over a 441-word vocabulary plus counts 1-30) and repetition counting (single most-confident class among 1-30). The authors report 76.5% exercise-detection 'accuracy' and 85.3% off-by-one repetition-counting accuracy on Olympia, along with fine-tuning, few-shot, and zero-shot experiments on Fit3D, RepCount, KIMORE, and UI-PRMD.

Significance. If the headline results were correctly measured, this would be a practically valuable contribution: it would demonstrate a single efficient transformer performing both exercise identification and rep counting at a scale (hundreds to thousands of exercises) well beyond prior work, using only privacy-preserving skeletal data from RGB video. The paper also deserves credit for assembling a large multi-source dataset and for reporting external generalization experiments, an OBO/MAE evaluation for counting, and an ablation comparing MotionBERT with Mediapipe keypoints. However, the abstract's central accuracy claim rests on a metric that is recall without false-positive penalty, and this undermines the current presentation of every detection result, including the zero-shot numbers.

major comments (4)
  1. [Section 4 (exercise detection scoring)] The central quantitative claim in the abstract is not supported by the metric. The 'exercise detection accuracy' is defined as Accuracy_i = |P_i ∩ G_i| / |G_i|, i.e., recall against the ground-truth word set, with no penalty for predicted words outside G_i and no requirement that P_i be non-empty or that the top prediction be correct. A trivial predictor that outputs all vocabulary words on every video receives 100%, and even a predictor that always outputs the single frequent word 'squat' receives full credit on every ground-truth label containing that word. The threshold τ=0.05 does not prevent this. The paper must report precision, F1, or exact-match accuracy, and must compare against a majority/frequent-label baseline; the 76.5% number should be withdrawn or relabeled as recall.
  2. [Section 4, Eq. (2); Fig. 6c-d; Table 3] The repetition-counting claim is not yet load-bearing because no trivial-predictor baseline is reported. Figure 6c-d shows heavy concentration at 5 and 10 repetitions; predicting the majority class (or the class-conditional mean) could plausibly achieve high off-by-one accuracy. The authors' own footnote 3 shows that fine-tuning on UI-PRMD makes the model 'always predict 10,' illustrating exactly the collapse that a baseline would expose. Please report majority-class, mean-count, and length-normalized baselines for OBO and MAE.
  3. [Abstract, Section 3.4/3.5, Table 1] The scale claim is internally inconsistent. The abstract states the dataset covers 'more than 1,900 exercises,' Table 1 lists Olympia as 1,306, Section 3.4 says the vocabulary represents '1,749 exercise labels,' and Section 3.5 describes 7,618 videos. The exercise count is a central contribution of the paper, so the authors need to reconcile these numbers and state clearly whether the unit is unique full labels, unique exercise names, or something else.
  4. [Section 4.2 and Table 3] The zero-shot/few-shot detection scores inherit the same recall-only definition and are evaluated against a vocabulary that was not extended to the target datasets' labels. For KIMORE and UI-PRMD, the reported scores therefore measure partial word overlap with an unadapted vocabulary rather than the ability to identify unseen exercises. Please report per-class precision and exact-match accuracy on these datasets as well.
minor comments (5)
  1. [Section 4.1] A paragraph beginning 'We examine the accuracy of the model on exercise detection and repetition counting' is repeated verbatim in the text.
  2. [Section 3.3 vs. Supplemental Information] The weight for the 'Uncategorized' category is inconsistent: Section 3.3 and Table 2 give 0.0, while the Supplemental Information lists 0.1.
  3. [Section 3.5 vs. Supplemental Information] The number of warmup exercises is inconsistent: Section 3.5 says 17 warmupX exercises, while the Supplemental vocabulary lists warmup1 through warmup19.
  4. [Section 4] The equations for OBO and MAE are unnumbered; please number them for reproducibility.
  5. [General] There is no data or code availability statement; given that Olympia is assembled from public and new sources, please state what will be released.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity found: LiFT's results are empirical evaluations on held-out splits and external benchmarks, not derivations from fitted inputs or self-citation.

full rationale

The paper's central claims are empirical: a model is trained on Olympia and evaluated on a random 80/10/10 split, plus fine-tune, few-shot, and zero-shot tests on Fit3D, RepCount, UI-PRMD, and KIMORE. The exercise-detection metric (Accuracy_i = |P_i ∩ G_i| / |G_i|, with P_i = {w_j : ŷ_{i,j} ≥ τ}) is a recall-style partial-credit score with no false-positive penalty; this is a metric-validity and evaluation-transparency concern, not circularity, because the reported numbers are computed from actual model outputs rather than constructed from the labels or fitted parameters. The repetition-counting OBO metric is standard in the cited RepCount/TransRAC literature. The dataset includes the authors' prior Variations dataset and the CHASE-2023 dataset (refs [33,34]), but those are data contributions and prior methods, not load-bearing uniqueness theorems or ansatz sources; the architecture is built on externally pretrained ViLT [51] and MotionBERT [24]. The single-class 'baseline' in Sec. 4.3.1 is an ablation of the model's own head, not a fitted input renamed as a prediction. The qualitative admission of 'entirely false classifications' (Sec. 4.1) is a limitation that strengthens metric-validity concerns, but no step in the derivation chain equates an output to an input by construction, and no self-citation carries the central claim.

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

The central claims rest on the quality of MotionBERT keypoints, the transferability of ViLT image-text pretraining to skeleton motion-text, and the validity of the hand-designed word weights and the recall-only evaluation metric. No new physical entities are introduced.

free parameters (5)
  • confidence threshold tau = 0.05
    Only predicted classes with softmax probability >= 0.05 count toward detection accuracy; chosen in Section 4, not justified.
  • word category weights W = {0, 0.1, 0.4, 0.5, 1}
    Hand-assigned weights in Section 3.3 affect the multi-hot loss and therefore what the model learns; same weights define the categories.
  • temporal smoothing window w = 3
    Smoothing window in Section 3.2; default choice without ablation.
  • interpolation length M = 64
    Per-chain interpolation to 64 points in Section 3.2; arbitrary resolution choice.
  • maximum input size = 640x384
    ViLT's maximum image size; forces temporal under-sampling and caps video length at 51.2 seconds.
assumptions (5)
  • domain assumption MotionBERT 3D pose estimates are accurate enough for exercise analysis
    Used in Section 3.1/3.5 as the only pose source; no error analysis on downstream tasks.
  • domain assumption ViLT pretrained on image-text pairs transfers to skeleton motion-text
    Section 3.4 initializes from ViLT; no ablation against training from scratch.
  • domain assumption Interpolated kinematic chains preserve task-relevant motion information
    Section 3.2 converts skeletons to images via 1D interpolation; no alternative representations tested.
  • ad hoc to paper The multi-hot weighted loss with hand-set weights is a valid objective
    Section 3.3 assigns weights by word category; these weights are specific to this paper and not derived from any principle.
  • domain assumption Olympia labels and repetition counts are correct
    Section 3.5 relies on third-party dataset labels and self-annotated open-source videos.

how reviews work

0 comments
Cite this review

Pith. "Pith review of (LiFT) Lightweight Fitness Transformer: A language-vision model for Remote Monitoring of Physical Training." pith.science (2026). https://pith.science/paper/HQMHQWF4

@misc{pith2026250606480,
  author       = {Pith},
  title        = {Pith review of: (LiFT) Lightweight Fitness Transformer: A language-vision model for Remote Monitoring of Physical Training},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/HQMHQWF4}},
  note         = {Machine review of arXiv:2506.06480}
}
read the original abstract

We introduce a fitness tracking system that enables remote monitoring for exercises using only a RGB smartphone camera, making fitness tracking more private, scalable, and cost effective. Although prior work explored automated exercise supervision, existing models are either too limited in exercise variety or too complex for real-world deployment. Prior approaches typically focus on a small set of exercises and fail to generalize across diverse movements. In contrast, we develop a robust, multitask motion analysis model capable of performing exercise detection and repetition counting across hundreds of exercises, a scale far beyond previous methods. We overcome previous data limitations by assembling a large-scale fitness dataset, Olympia covering more than 1,900 exercises. To our knowledge, our vision-language model is the first that can perform multiple tasks on skeletal fitness data. On Olympia, our model can detect exercises with 76.5% accuracy and count repetitions with 85.3% off-by-one accuracy, using only RGB video. By presenting a single vision-language transformer model for both exercise identification and rep counting, we take a significant step toward democratizing AI-powered fitness tracking.

Figures

Figures reproduced from arXiv: 2506.06480 by the authors.

Figure 1
Figure 1. The overview of LiFT where RGB videos are converted to skeletal keypoints [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. (a) Keypoints in Human3.6M format as detected by MotionBERT; (b) Resulting [PITH_FULL_IMAGE:figures/full_fig_p007_2.png] view at source ↗
Figure 3
Figure 3. Comparing architecture types for motion-language understanding. a) shows [PITH_FULL_IMAGE:figures/full_fig_p009_3.png] view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: Training and Validation losses on the Olympia dataset. [PITH_FULL_IMAGE:figures/full_fig_p010_4.png]
Figure 5
Figure 5. Figure 5: a) multi-class labeling using a shared Vocabulary and weighing words with [PITH_FULL_IMAGE:figures/full_fig_p013_5.png]
Figure 6
Figure 6. Figure 6: a) Comparing the accuracy of the model across the various exercise label word [PITH_FULL_IMAGE:figures/full_fig_p016_6.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

56 extracted references · 56 canonical work pages

  1. [1]

    US health care spending by payer and health condition, 1996-2016,

    J. L. Dieleman, J. Cao, A. Chapin, C. Chen, Z. Li, A. Liu, C. Horst, A. Kaldjian, T. Matyasz, K. W. Scott, A. L. Bui, M. Campbell, H. C. 20 Duber, A. C. Dunn, A. D. Flaxman, C. Fitzmaurice, M. Naghavi, N. Sa- dat, P. Shieh, E. Squires, K. Yeung, and C. J. L. Murray, “US health care spending by payer and health condition, 1996-2016,” vol. 323, no. 9, pp. 863–884

    work page 1996

  2. [2]

    Overcoming overuse: Improving musculoskeletal health care,

    J. R. Zadro, S. D´ ecary, M. O’Keeffe, Z. A. Michaleff, and A. C. Traeger, “Overcoming overuse: Improving musculoskeletal health care,” vol. 50, no. 3, pp. 113–115. Publisher: Journal of Orthopaedic & Sports Physical Therapy

    work page

  3. [3]

    Do physical therapists follow evidence-based guidelines when managing musculoskeletal conditions? systematic review,

    J. Zadro, M. O’Keeffe, and C. Maher, “Do physical therapists follow evidence-based guidelines when managing musculoskeletal conditions? systematic review,” vol. 9, no. 10, p. e032329. Publisher: British Medical Journal Publishing Group Section: Rehabilitation medicine

    work page

  4. [4]

    Activation of au- tophagy in human skeletal muscle is dependent on exercise intensity and AMPK activation,

    C. Schwalm, C. Jamart, N. Benoit, D. Naslain, C. Pr´ emont, J. Pr´ evet, R. Van Thienen, L. Deldicque, and M. Francaux, “Activation of au- tophagy in human skeletal muscle is dependent on exercise intensity and AMPK activation,” vol. 29, no. 8, pp. 3515–3526

    work page

  5. [5]

    Physical activity in cancer prevention and survival: A systematic review,

    A. McTiernan, C. M. Friedenreich, P. T. Katzmarzyk, K. E. Powell, R. Macko, D. Buchner, L. S. Pescatello, B. Bloodgood, B. Tennant, A. Vaux-Bjerke, S. M. George, R. P. Troiano, and K. L. Piercy, “Physical activity in cancer prevention and survival: A systematic review,” vol. 51, no. 6, pp. 1252–1261

    work page

  6. [6]

    Muscle mass, BMI, and mortality among adults in the united states: A population-based cohort study,

    M. K. Abramowitz, C. B. Hall, A. Amodu, D. Sharma, L. Androga, and M. Hawkins, “Muscle mass, BMI, and mortality among adults in the united states: A population-based cohort study,” vol. 13, no. 4, p. e0194697

    work page

  7. [7]

    An app with remote support achieves better adherence to home exercise pro- grams than paper handouts in people with musculoskeletal conditions: a randomised trial,

    T. E. Lambert, L. A. Harvey, C. Avdalis, L. W. Chen, S. Jeyalingam, C. A. Pratt, H. J. Tatum, J. L. Bowden, and B. R. Lucas, “An app with remote support achieves better adherence to home exercise pro- grams than paper handouts in people with musculoskeletal conditions: a randomised trial,” vol. 63, no. 3, pp. 161–167

    work page

  8. [8]

    Quality of knee strengthening exercises performed at home deteriorates after one week,

    U. H. Mitchell, H. Lee, H. E. Dennis, and M. K. Seeley, “Quality of knee strengthening exercises performed at home deteriorates after one week,” vol. 23, no. 1, p. 164. 21

    work page

Show all 56 references
  1. [9]

    Outcome measures for adherence to home exer- cises among patients with chronic low back pain: a systematic review,

    H. Uzawa and S. Davis, “Outcome measures for adherence to home exer- cises among patients with chronic low back pain: a systematic review,” vol. 30, no. 4, pp. 649–653

  2. [10]

    Mechanical diagnosis and therapy in back pain: Compliance and social cognitive theory,

    C. Hammer, L. Degerfeldt, and E. Denison, “Mechanical diagnosis and therapy in back pain: Compliance and social cognitive theory,” vol. 9, no. 4, pp. 190–197. Publisher: Taylor & Francis eprint: https://doi.org/10.1080/14038190701433241

    doi

  3. [11]

    Combined exercise and motivation program: Effect on the compliance and level of disability of patients with chronic low back pain: A ran- domized controlled trial,

    M. Friedrich, G. Gittler, Y. Halberstadt, T. Cermak, and I. Heiller, “Combined exercise and motivation program: Effect on the compliance and level of disability of patients with chronic low back pain: A ran- domized controlled trial,” vol. 79, no. 5, pp. 475–487

  4. [12]

    LiftRight: Quantifying strength training per- formance using a wearable sensor,

    S. Milanko and S. Jain, “LiftRight: Quantifying strength training per- formance using a wearable sensor,” vol. 16, p. 100115

  5. [13]

    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,” vol. 28, no. 2, pp. 468–477. Conference Name: IEEE Transactions on Neural Systems and Rehabil- itation Engineering

  6. [14]

    Skeleton- based human action evaluation using graph convolutional network for monitoring alzheimer’s progression,

    B. X. B. Yu, Y. Liu, K. C. C. Chan, Q. Yang, and X. Wang, “Skeleton- based human action evaluation using graph convolutional network for monitoring alzheimer’s progression,” vol. 119, p. 108095

  7. [15]

    Ef- ficient and robust skeleton-based quality assessment and abnormality detection in human action performance,

    A. Elkholy, M. E. Hussein, W. Gomaa, D. Damen, and E. Saba, “Ef- ficient and robust skeleton-based quality assessment and abnormality detection in human action performance,” vol. 24, no. 1, pp. 280–291. Conference Name: IEEE Journal of Biomedical and Health Informatics

  8. [16]

    LLMs are good action recognizers,

    H. Qu, Y. Cai, and J. Liu, “LLMs are good action recognizers,”

  9. [17]

    Live repetition counting,

    O. Levy and L. Wolf, “Live repetition counting,” in2015 IEEE Interna- tional Conference on Computer Vision (ICCV), pp. 3020–3028. ISSN: 2380-7504

  10. [18]

    Long action repetition counting with staged layer distillation,

    K. Chung, A. Kerrigan, and I. Dave, “Long action repetition counting with staged layer distillation,” 22

  11. [19]

    Repetitive activity counting by sight and sound

    Y. Zhang, L. Shao, and C. G. M. Snoek, “Repetitive activity counting by sight and sound.”

  12. [20]

    Context-aware and scale- insensitive temporal repetition counting,

    H. Zhang, X. Xu, G. Han, and S. He, “Context-aware and scale- insensitive temporal repetition counting,” in2020 IEEE/CVF Confer- ence on Computer Vision and Pattern Recognition (CVPR), pp. 667–

  13. [21]

    Mul- tiCounter: Multiple action agnostic repetition counting in untrimmed videos

    Y. Tang, W. Luo, J. Zhang, W. Huang, R. Jing, and D. Zhang, “Mul- tiCounter: Multiple action agnostic repetition counting in untrimmed videos.”

  14. [22]

    TransRAC: En- coding multi-scale temporal correlation with transformers for repetitive action counting,

    H. Hu, S. Dong, Y. Zhao, D. Lian, Z. Li, and S. Gao, “TransRAC: En- coding multi-scale temporal correlation with transformers for repetitive action counting,” pp. 19013–19022

  15. [23]

    BlazePose: On-device real-time body pose tracking

    V. Bazarevsky, I. Grishchenko, K. Raveendran, T. Zhu, F. Zhang, and M. Grundmann, “BlazePose: On-device real-time body pose tracking.”

  16. [24]

    MotionBERT: A unified perspective on learning human motion representations

    W. Zhu, X. Ma, Z. Liu, L. Liu, W. Wu, and Y. Wang, “MotionBERT: A unified perspective on learning human motion representations.”

  17. [25]

    A database of physical therapy exercises with variability of execution collected by wearable sensors,

    S. Garc ´ ıa-de Villa, A. Jim´ enez-Mart ´ ın, and J. J. Garc ´ ıa-Dom ´ ınguez, “A database of physical therapy exercises with variability of execution collected by wearable sensors,” vol. 9, p. 266

  18. [26]

    Pe- riodic physical activity information segmentation, counting and recog- nition from video,

    S.-H. Cheng, M. A. Sarwar, Y.-A. Daraghmi, T.-U. ˙Ik, and Y.-L. Li, “Pe- riodic physical activity information segmentation, counting and recog- nition from video,” vol. 11, pp. 23019–23031. Conference Name: IEEE Access

  19. [27]

    The KIMORE dataset: KInematic assess- ment of MOvement and clinical scores for remote monitoring of physical REhabilitation,

    M. Capecci, M. G. Ceravolo, F. Ferracuti, S. Iarlori, A. Monteri` u, L. Romeo, and F. Verdini, “The KIMORE dataset: KInematic assess- ment of MOvement and clinical scores for remote monitoring of physical REhabilitation,” vol. 27, no. 7, pp. 1436–1448

  20. [28]

    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,” vol. 3, no. 1, p. 2. Number: 1 Publisher: Multidisciplinary Digital Publishing Institute. 23

  21. [29]

    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,” in2021 IEEE/CVF Conference on Computer Vision and Pat- tern Recognition (CVPR), pp. 9914–9923, IEEE

  22. [30]

    Unlabeled action quality assessment based on multi-dimensional adap- tive constrained dynamic time warping

    “Unlabeled action quality assessment based on multi-dimensional adap- tive constrained dynamic time warping.”

  23. [31]

    AI human pose estimation: Yoga pose detection and correction,

    R. Gajbhiye, S. Jarag, P. Gaikwad, and S. Koparde, “AI human pose estimation: Yoga pose detection and correction,” vol. 7, no. 5

  24. [32]

    Counting out time: Class agnostic video repetition counting in the wild

    D. Dwibedi, Y. Aytar, J. Tompson, P. Sermanet, and A. Zisserman, “Counting out time: Class agnostic video repetition counting in the wild.”

  25. [33]

    PersonalPT: One- shot approach for skeletal-based repetitive action counting for physical therapy,

    A. Postlmayr, B. Garg, P. Cosman, and S. Dey, “PersonalPT: One- shot approach for skeletal-based repetitive action counting for physical therapy,” vol. 34, p. 100516

  26. [34]

    Short: Deep learning approach to skeletal performance evaluation of physical therapy exer- cises,

    B. Garg, A. Postlmayr, P. Cosman, and S. Dey, “Short: Deep learning approach to skeletal performance evaluation of physical therapy exer- cises,” inProceedings of the 8th ACM/IEEE International Conference on Connected Health: Applications, Systems and Engineering Technolo- gie...

  27. [35]

    Computer vision based on a mod- ular neural network for automatic assessment of physical therapy reha- bilitation activities,

    J. A. Francisco and P. S. Rodrigues, “Computer vision based on a mod- ular neural network for automatic assessment of physical therapy reha- bilitation activities,” vol. 31, pp. 2174–2183. Conference Name: IEEE Transactions on Neural Systems and Rehabilitation Engineering

  28. [36]

    Using learnable physics for real-time exercise form recommendations,

    A. Jaiswal, G. Chauhan, and N. Srivastava, “Using learnable physics for real-time exercise form recommendations,” inProceedings of the 17th ACM Conference on Recommender Systems, pp. 688–695

  29. [37]

    Online action representation using change detection and symbolic programming

    V. S. Nair, S. Sree, J. Joseph, and M. Sivaprakasam, “Online action representation using change detection and symbolic programming.”

  30. [38]

    MotionGPT: Human motion as a foreign language

    B. Jiang, X. Chen, W. Liu, J. Yu, G. Yu, and T. Chen, “MotionGPT: Human motion as a foreign language.” 24

  31. [39]

    MotionLLM: Understanding human behaviors from human motions and videos

    L.-H. Chen, S. Lu, A. Zeng, H. Zhang, B. Wang, R. Zhang, and L. Zhang, “MotionLLM: Understanding human behaviors from human motions and videos.”

  32. [40]

    AvatarGPT: All-in-one framework for motion understanding planning generation and beyond,

    Z. Zhou, Y. Wan, and B. Wang, “AvatarGPT: All-in-one framework for motion understanding planning generation and beyond,” pp. 1357–1366

  33. [41]

    Exploring the limits of transfer learning with a unified text-to-text transformer

    C. Raffel, N. Shazeer, A. Roberts, K. Lee, S. Narang, M. Matena, Y. Zhou, W. Li, and P. J. Liu, “Exploring the limits of transfer learning with a unified text-to-text transformer.”

  34. [42]

    Vicuna: An open-source chatbot impressing GPT-4 with 90%* Chat- GPT quality|LMSYS org

    “Vicuna: An open-source chatbot impressing GPT-4 with 90%* Chat- GPT quality|LMSYS org.”

  35. [43]

    LLaMA: Open and efficient foun- dation language models

    H. Touvron, T. Lavril, G. Izacard, X. Martinet, M.-A. Lachaux, T. Lacroix, B. Rozi` ere, N. Goyal, E. Hambro, F. Azhar, A. Rodriguez, A. Joulin, E. Grave, and G. Lample, “LLaMA: Open and efficient foun- dation language models.”

  36. [44]

    Neural discrete repre- sentation learning

    A. v. d. Oord, O. Vinyals, and K. Kavukcuoglu, “Neural discrete repre- sentation learning.”

  37. [45]

    Efficient estimation of word representations in vector space

    T. Mikolov, K. Chen, G. Corrado, and J. Dean, “Efficient estimation of word representations in vector space.”

  38. [46]

    Neural machine translation of rare words with subword units

    R. Sennrich, B. Haddow, and A. Birch, “Neural machine translation of rare words with subword units.”

  39. [47]

    Fast Word- Piece tokenization

    X. Song, A. Salcianu, Y. Song, D. Dopson, and D. Zhou, “Fast Word- Piece tokenization.”

  40. [48]

    EricGuo5513/HumanML3d

    C. Guo, “EricGuo5513/HumanML3d.” original-date: 2022-02- 26T21:02:58Z

    2022

  41. [49]

    Exploring vision transformers for 3d human motion-language models with motion patches,

    Q. Yu, M. Tanaka, and K. Fujiwara, “Exploring vision transformers for 3d human motion-language models with motion patches,” in2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pp. 937–946, IEEE

  42. [50]

    Human3.6m: Large scale datasets and predictive methods for 3d human sensing in 25 natural environments,

    C. Ionescu, D. Papava, V. Olaru, and C. Sminchisescu, “Human3.6m: Large scale datasets and predictive methods for 3d human sensing in 25 natural environments,” vol. 36, no. 7, pp. 1325–1339. Conference Name: IEEE Transactions on Pattern Analysis and Machine Intelligence

  43. [51]

    ViLT: Vision-and-language transformer without convolution or region supervision

    W. Kim, B. Son, and I. Kim, “ViLT: Vision-and-language transformer without convolution or region supervision.”

  44. [52]

    Papers with code - InfiniteRep dataset

    “Papers with code - InfiniteRep dataset.”

  45. [53]

    MM-fit - the dataset of full-body physical exercises captured by multi- ple time-synchronized wearable sensing devices

    “MM-fit - the dataset of full-body physical exercises captured by multi- ple time-synchronized wearable sensing devices.”

  46. [54]

    UCO physical reha- bilitation: New dataset and study of human pose estimation methods on physical rehabilitation exercises,

    R. Aguilar-Ortega, R. Berral-Soler, I. Jim´ enez-Velasco, F. J. Romero- Ram ´ ırez, M. Garc ´ ıa-Mar ´ ın, J. Zafra-Palma, R. Mu˜ noz-Salinas, R. Medina-Carnicer, and M. J. Mar ´ ın-Jim´ enez, “UCO physical reha- bilitation: New dataset and study of human pose estimation metho...

  47. [55]

    Bleu: a method for automatic evaluation of machine translation,

    K. Papineni, S. Roukos, T. Ward, and W.-J. Zhu, “Bleu: a method for automatic evaluation of machine translation,” inProceedings of the 40th Annual Meeting of the Association for Computational Linguistics (P. Isabelle, E. Charniak, and D. Lin, eds.), pp. 311–318, Association fo...

  48. [56]

    Supplemental Information ”Core Action”: ”weight”: 1.0, ”words”: [”squat”, ”lunge”, ”press”, ”push”, ”pull”, ”row”, ”deadlift”, ”hinge”, ”raise”, ”jump”, ”hop”, ”fly”, ”thruster”, ”step”, ”sit”, ”kneel”, ”kick”, ”swing”, ”scaptions”, ”curtsy”, ”plank”, ”crunch”, ”tap”, ”jack”, ...

Pith tools

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