Pith. sign in

REVIEW 3 major objections 5 minor 1 cited by

AthleticsPose: Authentic Sports Motion Dataset on Athletic Field and Evaluation of Monocular 3D Pose Estimation Ability

T0 review · 3 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read The paper claims that training a monocular 3D pose estimator on authentic, field-captured athletics motions instead of imitated sports motions reduces mean per-joint position error by about 75 percent.

desk verdict Genuinely useful dataset, but the headline 75% MPJPE claim rests on an uncontrolled baseline comparison and should be withdrawn or reframed. read the letter →

arxiv 2507.12905 v1 pith:H47L2OBF submitted 2025-07-17 cs.CV

Tomohiro Suzuki , Ryota Tanaka , Calvin Yeung , Keisuke Fujii This is my paper · ORCID
classification cs.CV
keywords AthleticsPosemonocular3DestimationsportsmotiondatasetmarkerlesscaptureMPJPEdomaingapinkinematicvalidationbiomechanics
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

The paper introduces AthleticsPose, a public dataset of roughly 500,000 frames of eight track-and-field events performed by 23 athletes on an outdoor field, with 3D joint coordinates captured by eight synchronized cameras and a markerless optical system. Its central claim is that training a monocular 3D pose estimator on these authentic motions instead of an imitated sports-motion dataset cuts mean per-joint position error (MPJPE) by about 75 percent: 29.45 mm versus 115.91 mm under ground-truth 2D input. The authors read this as evidence that the domain gap between scripted, lab-style imitated sports motions and real competitive movement is large and cannot be closed by model architecture alone. They also show that error grows for small on-screen subjects and front/back camera views, and that derived kinematic metrics split into two regimes: knee angles of the supporting leg retain subject-level distinctions, while maximum knee-drive velocity is systematically overestimated and blurs differences between athletes. The practical stakes are that monocular video analysis could become a usable, low-cost tool for field-based performance diagnostics, but only if the training data is authentic and the camera is placed to keep subjects large and near side-on.

What carries the argument

The load-bearing comparison is between two supervised training corpora: AthleticsPose, built from eight synchronized high-speed cameras and Theia3D markerless optical motion capture on an outdoor track-and-field setting, and AthletePose3D, an existing benchmark of imitated athletic movements captured in a lab. The model under test is MotionAGFormer, a transformer-GCN hybrid that lifts sequences of 81 frames of 2D keypoints to 3D joint coordinates, with ViTPose as the upstream 2D detector. The experimental design isolates the effect of training data by holding the 3D architecture fixed and comparing ground-truth-input and detector-input conditions. The supporting kinematic analysis computes knee angles of the supporting leg and maximum knee-drive velocity from predicted joints, using pelvis-relative, normalized coordinates. This setup lets the paper attribute the MPJPE difference to the realism of the motion content rather than to model capacity.

What would settle it

Capture a few of the same athletics trials simultaneously with Theia3D and with marker-based motion capture, and compare per-joint 3D errors during the fastest phases such as footstrike during sprinting and wrist release during throws. If markerless errors at wrists and ankles are on the order of the 75 percent gap, or show a consistent smoothing bias exactly where the model overestimates knee-drive velocity, then the dataset's ground truth, not the trained model, would be the main source of the reported results.

Watch

Extended reading notes

Core claim

The central result is empirical: a MotionAGFormer 2D-to-3D lifter trained on AthleticsPose ground-truth 2D keypoints reaches 29.45 mm MPJPE on authentic field motions, while the same architecture trained on AthletePose3D, a dataset of imitated athletic motions, reaches 115.91 mm under identical ground-truth input conditions. Even when the AthleticsPose-trained model receives 2D detections from a real 2D estimator instead of ground-truth keypoints, it stays at 41.99 mm, still roughly 65 percent better than the imitated-motion baseline. The paper attributes the gap to real versus imitated motion content: the imitated dataset lacks the speed, acceleration, and action-specific dynamics of actual competition, so a model trained on it does not transfer to real-world athletics. The same dataset analysis shows AthleticsPose contains substantially faster wrist, ankle, and hip motion than existing benchmarks, which is precisely the regime where the model is tested. The authors conclude that authentic field-captured data, not just sports-themed data, is what makes monocular 3D pose estimation useful for athletic performance analysis.

Load-bearing premise

The whole evaluation assumes the Theia3D markerless optical motion capture, after only internal bone-length consistency checks, gives joint coordinates accurate enough to serve as ground truth; if the system systematically biases joint positions during fast or occluded movement, every MPJPE, knee-angle, and knee-drive-velocity number measures agreement with that biased source rather than with true human motion.

Editorial extensions

If this is right

  • Training on authentic outdoor sports motion, rather than imitated lab motion, reduces MPJPE by roughly 75 percent in the same architecture, so future sports pose datasets should prioritize real competitive movement.
  • Monocular 3D pose estimation is usable for some kinematic indicators such as supporting-leg knee angle, which had an RMSE of 6.43 degrees, but not yet for velocity-based metrics such as knee-drive velocity, which had an RMSE of 0.77 m/s with consistent overestimation.
  • Camera placement matters: side views achieve 31.42 mm MPJPE versus 36.34 mm for front/back views, and large on-screen subjects achieve 33.27 mm versus 42.63 mm for small ones.
  • Models trained on AthleticsPose also beat a model trained on Human3.6M by about 70 percent, indicating that general daily-activity data is a weak surrogate for sports motion.
  • Per-joint errors concentrate in wrists and elbows during throwing events, so upper-limb kinematic analysis from monocular video is not yet reliable for discus and shot put.

Reading between the lines

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

  • The paper does not test whether Theia3D's markerless labels carry a systematic bias; if they do, part of the 75 percent gap could reflect shared bias between training labels and evaluation labels rather than true motion accuracy, and a small marker-based validation study on the same trials would settle this.
  • The dataset makes it possible to quantify when imitated motion fails by reweighting or augmenting imitated data to match AthleticsPose's speed and acceleration distributions, which would isolate how much of the domain gap is attributable to motion dynamics alone.
  • The camera-view and scale results suggest a practical field-deployment rule: keep the athlete's bounding box above a size threshold and prefer side-on camera positions for running events.
  • Because knee-drive velocity is consistently overestimated, a per-subject calibration curve fitted to prediction errors might recover the relational differences between athletes, a test the paper does not run.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

3 major / 5 minor

Summary. The paper introduces AthleticsPose, a public dataset of authentic athletics motions captured from 23 athletes on an outdoor field using eight synchronized cameras and a Theia3D markerless motion-capture system. The authors fine-tune ViTPose for 2D pose estimation and train MotionAGFormer for 3D pose estimation under three input conditions (GT, DET pre-trained, DET fine-tuned), then evaluate MPJPE/P-MPJPE by action, joint, camera view, and input scale. They also perform case studies of sprint kinematics (supporting-leg knee angle and knee-drive velocity). The headline claim, stated in the abstract and Section 5.4, is that a model trained on AthleticsPose reduces MPJPE by about 75% compared with a baseline trained on the imitated-motion dataset AthletePose3D (115.91 mm vs. 29.45 mm).

Significance. If the dataset labels are accurate and the comparison is properly controlled, AthleticsPose would be a valuable contribution: it is the first openly available 3D pose dataset of authentic athletics motions captured outdoors, it ships code and checkpoints, and it goes beyond standard error metrics by examining kinematic indicators and their reliability. The honest reporting of limitations, such as the failure to recover knee-drive velocity differences, is a strength. However, the central 75% claim is not yet supported because the comparison against AthletePose3D is confounded, and the markerless ground-truth labels lack independent validation. These issues affect the validity of every quantitative result in the paper, so they must be addressed before the main claims can be accepted.

major comments (3)
  1. [Section 5.4, Table 2] The central claim that AthleticsPose training reduces MPJPE by ~75% relative to an imitated-motion baseline is not supported by a controlled experiment. As stated in Section 5.2.2, the AthletePose3D baseline is the authors' existing public checkpoint and was not retrained with the AthleticsPose pipeline. The comparison therefore conflates 'authentic vs. imitated' with multiple other differences: laboratory vs. outdoor environment, 4 vs. 23 training subjects, differing action coverage, and differing training protocols (frame length, normalization, epoch count, 2D-input format). The sentence in Section 5.4 that attributes the improvement to 'Realistic motions in our dataset as opposed to Imitated motions' is not justified by this experimental design. To support the causal claim, the authors should retrain a model on AthletePose3D data under matched conditions (same 81-frame input, same absolute-value normalization, same 100-epoch schedule, same 17-keypoint format) and evaluate on the same test split, or substantially weaken the abstract and conclusion claims.
  2. [Section 3.2] The ground-truth labels produced by Theia3D are validated only through internal bone-length consistency checks; no comparison to marker-based motion capture or to an independent measurement modality is reported. Since these labels serve as both training supervision and evaluation ground truth, any systematic markerless bias (for example, smoothing of high-speed footstrikes or misplacement of wrist joints during throws) would propagate into every MPJPE, P-MPJPE, knee-angle, and knee-drive-velocity result, and would also undermine the claim that the dataset captures authentic athletic motion. The paper should include a validation study on a subset of trials (e.g., simultaneous marker-based capture or an alternate multi-view 3D reconstruction) and report per-joint error characteristics for Theia3D in the present capture conditions.
  3. [Section 5.4, Figures 4-5 (with Section 3.5)] The kinematic case-study results are presented as demonstrating that the model 'captures subject-specific patterns' (ANOVA and Tukey HSD), but the same paragraph reports that predicted angles underestimate inter-subject differences (14.66° vs. 24.07°) and create a distinction not present in the ground truth (S21/S22, p=0.089 vs. p=0.003). The current wording overstates the positive finding and underweights the bias. This matters because the utility for sports analysis depends on faithfully reproducing relational differences, not merely on detecting some difference. The text should be revised to state explicitly that the model produces biased effect sizes and can create false positives in statistical comparisons, and the abstract's emphasis on 'potential to capture individual differences' should be qualified accordingly.
minor comments (5)
  1. [Section 5.4] The text says the Human3.6M baseline P-MPJPE decreased from 98.90 mm to 31.40 mm, but Table 2 reports the baseline P-MPJPE as 98.80 mm; please correct the inconsistency.
  2. [Section 4.1 / Table 1] The abstract says the dataset comprises 'approximately 500,000 frames,' while Table 1 lists 62.5K poses and 500K frames; please clarify that 500K refers to multi-view video frames (8 cameras × 62.5K 3D pose instances).
  3. [Section 5.4] The note 'Please note that all results presented below are evaluation results using AthleticsPose (DET fine-tuned), unless otherwise mentioned' appears after Table 3, leaving Table 3 ambiguous; move the note before Table 3 or state the configuration in the table caption.
  4. [Section 5.4] In the knee-angle analysis, the sentence 'An analysis of variance (ANOVA) and subsequent Tukey's HSD post-hoc test confirmed that 3D pose estimation successfully captures subject-specific patterns' is too strong in light of the biases reported later in the same paragraph; the wording should be more neutral, e.g., 'captures gross subject-level differences but with attenuated magnitudes and potential false distinctions.'
  5. [Section 3.5 / Section 5.3.3] The description 'the distance between the knees is calculated' for knee-drive velocity is ambiguous; please specify whether this is the 3D Euclidean distance or a directional component (e.g., in the sagittal plane) and define how the velocity is computed and smoothed.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity found: the central comparison is empirical and not definitionally or statistically forced.

full rationale

The paper's central claim—that a model trained on AthleticsPose outperforms one trained on AthletePose3D—is an empirical comparison of independently trained checkpoints reported in Table 2, not a quantity derived from a fitted parameter or from an equation whose output is its own input. The reduction in MPJPE from 115.91 mm to 29.45 mm is a reported measurement, not a prediction manufactured from a fit. The use of AthletePose3D, a prior work by overlapping authors, provides a public baseline checkpoint; the load here is the empirical result, not an invocation of an unverified uniqueness theorem or ansatz. The shared Theia3D markerless pipeline for both training labels and evaluation labels is a validity concern about absolute ground-truth accuracy, but it does not make the derivation circular: the model output is not defined in terms of the evaluation labels, and the training and test splits are by subject. The comparison is confounded by environment, action coverage, and training protocol, but confounding is a correctness risk, not a circularity identified by the required reduction evidence. Therefore, no specific circular step meeting the hard-rule standard can be quoted or exhibited, and the appropriate finding is no significant circularity (score 0).

Assumptions & free parameters 1 free parameters · 3 assumptions · 0 invented entities

The paper is an empirical benchmark, not a derivation. It introduces no new theoretical entities or fitted constants; the main hand-chosen analysis parameters are the relative-scale thresholds used to bin inputs into Large, Medium, and Small. The central assumptions are that the markerless mocap system produces valid ground truth, that projected 2D keypoints are valid supervision, and that the AthletePose3D baseline is comparable.

free parameters (1)
  • Relative scale thresholds for input scale bins = 0.33 and 0.67 of the maximum bounding-box height
    Hand-chosen cutoffs define Large, Medium, and Small categories in Table 5. The monotonic error trend is likely robust to small changes, but the exact per-bin MPJPE values and the 'performance degrades with smaller scale' numbers depend on these cutoffs.
assumptions (3)
  • domain assumption Theia3D markerless motion capture provides sufficiently accurate 3D joint positions to serve as ground truth.
    Section 3.1 uses Theia3D as the motion capture system and Section 3.2 only applies bone-length consistency checks. No comparison to marker-based motion capture is reported, so systematic errors in Theia3D would affect both training labels and evaluation.
  • domain assumption Projecting 3D markerless joints onto 2D images yields valid keypoint supervision for ViTPose.
    Section 3.2 states that validated 3D joint positions were projected onto images using calibrated intrinsic and extrinsic parameters. This assumes the calibration and the markerless joint positions are accurate enough for supervised 2D detection.
  • domain assumption The AthletePose3D checkpoint and the AthleticsPose-trained model are comparable except for the training data.
    Section 5.2.2 uses the public AthletePose3D model as the imitated baseline without retraining under identical schedules or hyperparameters, yet Section 5.4 attributes the performance gap to 'Realistic' versus 'Imitated' motions.

how reviews work

0 comments
Cite this review

Pith. "Pith review of AthleticsPose: Authentic Sports Motion Dataset on Athletic Field and Evaluation of Monocular 3D Pose Estimation Ability." pith.science (2026). https://pith.science/paper/H47L2OBF

@misc{pith2026250712905,
  author       = {Pith},
  title        = {Pith review of: AthleticsPose: Authentic Sports Motion Dataset on Athletic Field and Evaluation of Monocular 3D Pose Estimation Ability},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/H47L2OBF}},
  note         = {Machine review of arXiv:2507.12905}
}
read the original abstract

Monocular 3D pose estimation is a promising, flexible alternative to costly motion capture systems for sports analysis. However, its practical application is hindered by two factors: a lack of realistic sports datasets and unclear reliability for sports tasks. To address these challenges, we introduce the AthleticsPose dataset, a new public dataset featuring ``real'' motions captured from 23 athletes performing various athletics events on an athletic field. Using this dataset, we trained a representative 3D pose estimation model and performed a comprehensive evaluation. Our results show that the model trained on AthleticsPose significantly outperforms a baseline model trained on an imitated sports motion dataset, reducing MPJPE by approximately 75 %. These results show the importance of training on authentic sports motion data, as models based on imitated motions do not effectively transfer to real-world motions. Further analysis reveals that estimation accuracy is sensitive to camera view and subject scale. In case studies of kinematic indicators, the model demonstrated the potential to capture individual differences in knee angles but struggled with higher-speed metrics, such as knee-drive velocity, due to prediction biases. This work provides the research community with a valuable dataset and clarifies the potential and practical limitations of using monocular 3D pose estimation for sports motion analysis. Our dataset, code, and checkpoints are available at https://github.com/SZucchini/AthleticsPose.

Figures

Figures reproduced from arXiv: 2507.12905 by the authors.

Figure 1
Figure 1. An overview of AthleticsPose dataset, containing authentic motions on an athletic field, with example camera views. [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 3
Figure 3. Comparison of Cumulative distribution of joint [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figure 4
Figure 4. Comparison of GT and predicted knee angles by [PITH_FULL_IMAGE:figures/full_fig_p008_4.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. VIFSS: View-Invariant and Figure Skating-Specific Pose Representation Learning for Temporal Action Segmentation

    cs.CV 2025-08 unverdicted novelty 6.0 of 10

    Claims a view-invariant pose representation method for figure skating jump segmentation reaches over 92% F1@50, but the submitted text is a different manuscript, leaving the claim unverified.

Reference graph

Works this paper leans on

44 extracted references · 33 canonical work pages · cited by 1 Pith paper

  1. [1]

    Z. Cao, G. Hidalgo Martinez, T. Simon, S. Wei, and Y. A. Sheikh. 2019. OpenPose: Realtime Multi-Person 2D Pose Estimation using Part Affinity Fields. IEEE Transactions on Pattern Analysis and Machine Intelligence (2019)

    work page 2019

  2. [2]

    Bowen Cheng, Bin Xiao, Jingdong Wang, Honghui Shi, Thomas S Huang, and Lei Zhang. 2020. Higherhrnet: Scale-aware representation learning for bottom-up human pose estimation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition . 5386–5395

    work page 2020

  3. [3]

    Hai Ci, Chunyu Wang, Xiaoxuan Ma, and Yizhou Wang. 2019. Optimizing Network Structure for 3D Human Pose Estimation. In Proceedings of the IEEE International Conference on Computer Vision . 2262–2271

    work page 2019

  4. [4]

    MMPose Contributors. 2020. OpenMMLab Pose Estimation Toolbox and Bench- mark. https://github.com/open-mmlab/mmpose

    work page 2020

  5. [5]

    Takashi Fukushima, Patrick Blauberger, Tiago Guedes Russomanno, and Martin Lames. 2024. The potential of human pose estimation for motion capture in sports: a validation study. Sports Engineering 27, 1 (2024), 19

    work page 2024

  6. [6]

    Zigang Geng, Ke Sun, Bin Xiao, Zhaoxiang Zhang, and Jingdong Wang. 2021. Bottom-up human pose estimation via disentangled keypoint regression. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition . 14676–14686

    work page 2021

  7. [7]

    Lucas D Haberkamp, Micah C Garcia, and David M Bazett-Jones. 2022. Validity of an artificial intelligence, human pose estimation model for measuring single-leg squat kinematics. Journal of Biomechanics 144 (2022), 111333

    work page 2022

  8. [8]

    Rebecca I Hamilton, Zornitza Glavcheva-Laleva, Md Imdadul Haque Milon, Yesh- win Anil, Jenny Williams, Peter Bishop, and Catherine Holt. 2024. Comparison of computational pose estimation models for joint angles with 3D motion capture. Journal of Bodywork and Movement Therapies 40 (2024), 315–319

    work page 2024

Show all 44 references
  1. [9]

    Christian Keilstrup Ingwersen, Janus Nørtoft Jensen, Morten Rieger Hannemose, and Anders Bjorholm Dahl. 2023. Evaluating current state of monocular 3D pose models for golf. In Northern Lights Deep Learning Workshop 2023. Septentrio Academic Publishing

    2023

  2. [10]

    Christian Keilstrup Ingwersen, Christian Møller Mikkelstrup, Janus Nørtoft Jensen, Morten Rieger Hannemose, and Anders Bjorholm Dahl. 2023. Sportspose- a dynamic 3d sports pose dataset. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition . 5219–5228

    2023

  3. [11]

    Catalin Ionescu, Dragos Papava, Vlad Olaru, and Cristian Sminchisescu. 2013. Human3. 6m: Large scale datasets and predictive methods for 3d human sensing in natural environments. IEEE transactions on pattern analysis and machine AthleticsPose: Authentic Sports Motion Dataset o...

    2013

  4. [12]

    Tianjian Jiang, Johsan Billingham, Sebastian Müksch, Juan Zarate, Nicolas Evans, Martin R Oswald, Marc Polleyfeys, Otmar Hilliges, Manuel Kaufmann, and Jie Song. 2024. WorldPose: A world cup dataset for global 3D human pose estimation. In European Conference on Computer Vision...

    2024

  5. [13]

    Hanbyul Joo, Hao Liu, Lei Tan, Lin Gui, Bart Nabbe, Iain Matthews, Takeo Kanade, Shohei Nobuhara, and Yaser Sheikh. 2015. Panoptic studio: A massively multi- view system for social motion capture. In Proceedings of the IEEE International Conference on Computer Vision . 3334–3342

    2015

  6. [14]

    Ross, and Angjoo Kanazawa

    Ruilong Li, Sha Yang, David A. Ross, and Angjoo Kanazawa. 2021. AI Choreog- rapher: Music Conditioned 3D Dance Generation with AIST++. 2021 IEEE/CVF International Conference on Computer Vision (ICCV) (2021), 13381–13392. https: //api.semanticscholar.org/CorpusID:236882798

    2021

  7. [15]

    Wenhao Li, Hong Liu, Runwei Ding, Mengyuan Liu, Pichao Wang, and Wenming Yang. 2022. Exploiting temporal contexts with strided transformer for 3d human pose estimation. IEEE Transactions on Multimedia 25 (2022), 1282–1293

    2022

  8. [16]

    Wenhao Li, Hong Liu, Hao Tang, and Pichao Wang. 2023. Multi-Hypothesis Representation Learning for Transformer-Based 3D Human Pose Estimation. Pattern Recognition 141 (2023), 109631

    2023

  9. [17]

    Wenhao Li, Hong Liu, Hao Tang, Pichao Wang, and Luc Van Gool. 2022. MH- Former: Multi-Hypothesis Transformer for 3D Human Pose Estimation. In Pro- ceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). 13147–13156

    2022

  10. [18]

    Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dollár, and C Lawrence Zitnick. 2014. Microsoft coco: Common objects in context. In Computer Vision–ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proce...

    2014

  11. [19]

    Jiajie Liu, Mengyuan Liu, Hong Liu, and Wenhao Li. 2025. TCPFormer: Learning Temporal Correlation with Implicit Pose Proxy for 3D Human Pose Estimation. arXiv preprint arXiv:2501.01770 (2025)

    2025 arXiv

  12. [20]

    Dushyant Mehta, Helge Rhodin, Dan Casas, Pascal Fua, Oleksandr Sotnychenko, Weipeng Xu, and Christian Theobalt. 2017. Monocular 3d human pose estimation in the wild using improved cnn supervision. In 2017 international conference on 3D vision (3DV) . IEEE, 506–516

    2017

  13. [21]

    Dimitrios Menychtas, Nikolaos Petrou, Ioannis Kansizoglou, Erasmia Giannakou, Athanasios Grekidis, Antonios Gasteratos, Vassilios Gourgoulis, Eleni Douda, Ilias Smilios, Maria Michalopoulou, et al. 2023. Gait analysis comparison between manual marking, 2D pose estimation algor...

    2023

  14. [22]

    Aiden Nibali, Joshua Millward, Zhen He, and Stuart Morgan. 2021. ASPset: An Outdoor Sports Pose Video Dataset With 3D Keypoint Annotations. Image and Vision Computing (2021), 104196. doi:10.1016/j.imavis.2021.104196

    2021

  15. [23]

    Dario Pavllo, Christoph Feichtenhofer, David Grangier, and Michael Auli. 2019. 3D human pose estimation in video with temporal convolutions and semi-supervised training. In Conference on Computer Vision and Pattern Recognition (CVPR)

    2019

  16. [24]

    AJ Piergiovanni and Michael S Ryoo. 2019. Early detection of injuries in mlb pitchers from video. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops. 0–0

    2019

  17. [25]

    Wenkang Shan, Zhenhua Liu, Xinfeng Zhang, Zhao Wang, Kai Han, Shanshe Wang, Siwei Ma, and Wen Gao. 2023. Diffusion-Based 3D Human Pose Estimation with Multi-Hypothesis Aggregation. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) . 14761–14771

    2023

  18. [26]

    Leonid Sigal, Alexandru O Balan, and Michael J Black. 2010. Humaneva: Synchro- nized video and motion capture dataset and baseline algorithm for evaluation of articulated human motion. International journal of computer vision 87, 1 (2010), 4–27

    2010

  19. [27]

    Babak Taati Soroush Mehraban, Vida Adeli. 2024. MotionAGFormer: Enhanc- ing 3D Human Pose Estimation with a Transformer-GCNFormer Network. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision

    2024

  20. [28]

    Ke Sun, Bin Xiao, Dong Liu, and Jingdong Wang. 2019. Deep high-resolution representation learning for human pose estimation. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition . 5693–5703

    2019

  21. [29]

    Tomohiro Suzuki, Kazuya Takeda, and Keisuke Fujii. 2022. Automatic fault detection in race walking from a smartphone camera via fine-tuning pose es- timation. In 2022 IEEE 11th Global Conference on Consumer Electronics (GCCE) . IEEE, 631–632

    2022

  22. [30]

    Tomohiro Suzuki, Kazuya Takeda, and Keisuke Fujii. 2024. Automatic Detection of Faults in Simulated Race Walking from a Fixed Smartphone Camera.International Journal of Computer Science in Sport 23, 1 (2024), 22–36. doi:doi:10.2478/ijcss- 2024-0002

    2024 doi

  23. [31]

    Ryota Tanaka, Tomohiro Suzuki, and Keisuke Fujii. 2024. 3D Pose-Based Temporal Action Segmentation for Figure Skating: A Fine-Grained and Jump Procedure- Aware Annotation Approach. In Proceedings of the 7th ACM International Work- shop on Multimedia Content Analysis in Sports ...

    2024

  24. [32]

    Ryota Tanaka, Tomohiro Suzuki, Kazuya Takeda, and Keisuke Fujii. 2023. Auto- matic Edge Error Judgment in Figure Skating Using 3D Pose Estimation from a Monocular Camera and IMUs. In Proceedings of the 6th International Workshop on Multimedia Content Analysis in Sports . 41–48

    2023

  25. [33]

    Ryota Tanaka, Tomohiro Suzuki, Kazuya Takeda, and Keisuke Fujii. 2023. Auto- matic Edge Error Judgment in Figure Skating Using 3D Pose Estimation from Inertial Sensors. In 2023 IEEE 12th Global Conference on Consumer Electronics (GCCE). IEEE, 1099–1100

    2023

  26. [34]

    Petra Torvinen, Keijo S Ruotsalainen, Shuang Zhao, Neil Cronin, Olli Ohtonen, and Vesa Linnamo. 2024. Evaluation of 3D markerless motion capture system accuracy during skate skiing on a treadmill. Bioengineering 11, 2 (2024), 136

    2024

  27. [35]

    Matt Trumble, Andrew Gilbert, Charles Malleson, Adrian Hilton, and John Col- lomosse. 2017. Total Capture: 3D Human Pose Estimation Fusing Video and Inertial Sensors. In 2017 British Machine Vision Conference (BMVC)

    2017

  28. [36]

    Scott D Uhlrich, Antoine Falisse, Łukasz Kidziński, Julie Muccini, Michael Ko, Akshay S Chaudhari, Jennifer L Hicks, and Scott L Delp. 2023. OpenCap: Human movement dynamics from smartphone videos. PLoS computational biology 19, 10 (2023), e1011462

    2023

  29. [37]

    Timo Von Marcard, Roberto Henschel, Michael J Black, Bodo Rosenhahn, and Gerard Pons-Moll. 2018. Recovering accurate 3d human pose in the wild using imus and a moving camera. InProceedings of the European conference on computer vision (ECCV). 601–617

    2018

  30. [38]

    Edward P Washabaugh, Thanikai Adhithiyan Shanmugam, Rajiv Ranganathan, and Chandramouli Krishnan. 2022. Comparing the accuracy of open-source pose estimation methods for measuring gait kinematics. Gait & posture 97 (2022), 188–195

    2022

  31. [39]

    Yufei Xu, Jing Zhang, Qiming Zhang, and Dacheng Tao. 2022. Vitpose: Simple vision transformer baselines for human pose estimation. Advances in neural information processing systems 35 (2022), 38571–38584

    2022

  32. [40]

    Calvin Yeung, Kenjiro Ide, and Keisuke Fujii. 2024. Autosoccerpose: Automated 3d posture analysis of soccer shot movements. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition . 3214–3224

    2024

  33. [41]

    Calvin Yeung, Tomohiro Suzuki, Ryota Tanaka, Zhuoer Yin, and Keisuke Fujii

  34. [42]

    Jinlu Zhang, Zhigang Tu, Jianyu Yang, Yujin Chen, and Junsong Yuan. 2022. MixSTE: Seq2seq Mixed Spatio-Temporal Encoder for 3D Human Pose Estimation in Video. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). 13232–13242

    2022

  35. [43]

    Wentao Zhu, Xiaoxuan Ma, Zhaoyang Liu, Libin Liu, Wayne Wu, and Yizhou Wang. 2023. MotionBERT: A Unified Perspective on Learning Human Motion Rep- resentations. In Proceedings of the IEEE/CVF International Conference on Computer Vision

    2023

  36. [2025]

    InProceedings of the Computer Vision and Pattern Recognition Conference

    AthletePose3D: A benchmark dataset for 3D human pose estimation and kinematic validation in athletic movements. InProceedings of the Computer Vision and Pattern Recognition Conference . 5945–5956

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.