AthleticsPose: Authentic 3D Sports Pose Dataset

• AthleticsPose is a public dataset that provides authentic outdoor athletics motion captured with markerless, multi-camera systems, featuring 83 joints per frame and over 500K frames.
• The framework employs a robust four-stage annotation pipeline, standardizing 3D pose labels by converting data from an 83-joint representation to common formats like COCO and Human3.6M.
• Empirical findings reveal that training on authentic athletics data reduces MPJPE by nearly 75% compared to imitated-motion datasets, underscoring its practical impact for sports performance analysis.

Searching arXiv for AthleticsPose and closely related sports pose estimation work to ground the article in current literature. AthleticsPose is a public dataset and evaluation framework for monocular 3D human pose estimation in athletics, introduced to study authentic sports motion under realistic outdoor field conditions and to assess whether current monocular methods are reliable enough for sports performance analysis (Suzuki et al., 17 Jul 2025). It is defined by two linked contributions: a benchmark built from real athletics performances by university athletes on an athletic field, and an empirical study showing that training on authentic athletic motion substantially outperforms training on imitated or generic human-motion datasets for 3D pose estimation in this domain (Suzuki et al., 17 Jul 2025). In the broader sports-pose literature, AthleticsPose occupies a distinctive position by emphasizing authentic field-captured athletics motion rather than laboratory capture, imitated actions, or narrowly scoped single-discipline datasets, while also evaluating downstream kinematic quantities rather than restricting assessment to joint-position error alone (Suzuki et al., 17 Jul 2025).

1. Definition and scope

AthleticsPose is presented as the first open 3D pose dataset of authentic athletics motion captured on an athletic field (Suzuki et al., 17 Jul 2025). Its stated purpose is to address two barriers to practical monocular 3D pose analysis in sports: the scarcity of realistic sports datasets and the lack of clarity about whether pose estimates are reliable for actual sports-analysis tasks (Suzuki et al., 17 Jul 2025).

The dataset contains 23 athletes in the released benchmark split, although the full data collection involved 24 university student athletes aged 19–26 years, including 15 male and 9 female participants (Suzuki et al., 17 Jul 2025). The motions are drawn from eight athletics events: starting dash, sprinting, distance running, race walking, hurdling, shot put, discus throw, and javelin throw; javelin throw was excluded from evaluation because only one subject performed it (Suzuki et al., 17 Jul 2025). The released benchmark contains 83 joint points per frame, about 62.5K poses, about 500K frames, and recordings at 30/60 FPS from 8 cameras in an outdoor athletic-field environment (Suzuki et al., 17 Jul 2025).

A central claim of the work is that authentic sports motion matters decisively for model training. AthleticsPose is explicitly contrasted with datasets dominated by daily activities, controlled lab capture, or imitated sports actions, and the paper argues that such alternatives fail to reproduce the high speeds, high accelerations, and movement strategies of athletes performing their own specialty events on a real field (Suzuki et al., 17 Jul 2025). This framing places AthleticsPose in direct dialogue with sports-pose resources such as AthletePose3D (Yeung et al., 10 Mar 2025), SportsPose (Ingwersen et al., 2023), and sports-specific pose-estimation pipelines for athlete analysis (Einfalt et al., 2018, Einfalt et al., 2020).

2. Data acquisition and annotation pipeline

AthleticsPose was captured with the Theia3D optical markerless motion capture system using eight high-speed Miqus Video cameras from Qualisys Inc. at HD resolution (Suzuki et al., 17 Jul 2025). The system operated in a markerless setup, which the authors emphasize as important for preserving natural athlete motion (Suzuki et al., 17 Jul 2025). Temporal synchronization was achieved with hardware synchronization, and camera calibration was performed in Qualisys Track Manager using a wand-type calibration kit (Suzuki et al., 17 Jul 2025).

The post-processing pipeline is described as four sequential stages: identifying frame intervals where all 83 joints were detected, performing bone-length consistency checks to reject erroneous markerless reconstructions, clipping synchronized corresponding frames from all camera videos, and projecting validated 3D joint coordinates into each camera view using calibrated intrinsic and extrinsic parameters (Suzuki et al., 17 Jul 2025). The paper does not provide explicit projection equations or full calibration matrices, but it states that 3D joint coordinates were projected into 2D using calibrated camera parameters (Suzuki et al., 17 Jul 2025).

For model training, the 3D labels were transformed from the camera coordinate system to the pixel coordinate system, and both 2D inputs and 3D labels were converted to root-relative coordinates using the pelvis as root in the Human3.6M convention, then normalized to $[-1,1]$ (Suzuki et al., 17 Jul 2025). The format conversion also maps the original 83-joint representation to a 17-joint representation; in the practical 3D lifting setup, 2D detections in COCO format are converted to Human3.6M format before being passed to the 3D estimator (Suzuki et al., 17 Jul 2025).

A plausible implication is that AthleticsPose was designed to interoperate with standard monocular 2D-to-3D lifting pipelines rather than requiring a bespoke skeleton definition throughout the learning stack. This interoperability distinguishes it from some sports datasets that primarily function as capture resources rather than benchmarking substrates (Suzuki et al., 17 Jul 2025, Ingwersen et al., 2023).

3. Authenticity as a dataset principle

In AthleticsPose, “authentic” means that athletes perform their own specialty events on an actual outdoor athletics field under realistic performance conditions (Suzuki et al., 17 Jul 2025). The paper emphasizes that authenticity is not a rhetorical label but a measurable property of the kinematic distribution. Compared with AthletePose3D, AthleticsPose exhibits higher mean joint speeds at wrists, ankles, and hips, reported respectively as 3.69 m/s vs. 3.06 m/s, 3.18 m/s vs. 2.69 m/s, and 2.94 m/s vs. 2.31 m/s (Suzuki et al., 17 Jul 2025). For acceleration, AthleticsPose also shows more extreme values at wrists and ankles, reported as 46.48 m/s$^2$ vs. 36.53 m/s$^2$ and 32.88 m/s$^2$ vs. 21.46 m/s$^2$, while hip acceleration is described as comparable (Suzuki et al., 17 Jul 2025).

The paper further notes a rightward shift in cumulative speed and acceleration distributions. One example given is that around 50% of wrist speeds in AthletePose3D lie at or below 3.0 m/s, whereas only about 40% of AthleticsPose wrist speeds do (Suzuki et al., 17 Jul 2025). The authors interpret this as evidence that AthleticsPose better captures explosive, upper-limit human movement such as sprinting, hurdling, and throwing (Suzuki et al., 17 Jul 2025).

This emphasis on authenticity differentiates AthleticsPose from several neighboring resources. AthletePose3D is explicitly sports-specific and contains approximately 1.3 million frames across running, track and field, and figure skating, but it includes competition-like motion captured in lab or rink settings and only one track-and-field subject (Yeung et al., 10 Mar 2025). SportsPose is a markerless 3D sports dataset with dynamic motion and quantitative validation, but its activities are soccer, volleyball, jump, baseball pitch, and tennis rather than athletics-specific events (Ingwersen et al., 2023). AthleticsPose is therefore narrower in sports taxonomy than some general sports datasets but more focused on the event ecology of athletics (Suzuki et al., 17 Jul 2025).

4. Benchmark protocol and monocular 3D pipeline

The representative monocular 3D pipeline in AthleticsPose uses ViTPose for monocular 2D keypoint estimation and MotionAGFormer for 2D-to-3D lifting (Suzuki et al., 17 Jul 2025). ViTPose is used as a strong top-down 2D pose estimator, and MotionAGFormer is selected as a representative state-of-the-art lifting model that has performed well on both general and sports pose datasets (Suzuki et al., 17 Jul 2025). The implementation is in Python 3 / PyTorch, with ViTPose trained in MMPose (Suzuki et al., 17 Jul 2025).

For 2D pose training, the authors use ViTPose base, fine-tuned from COCO-pretrained weights because training from scratch was ineffective (Suzuki et al., 17 Jul 2025). Ground-truth 83 marker points are projected to each image, the largest enclosing rectangle is used as the target person bounding box, and the 83 points are converted to 17 COCO keypoints (Suzuki et al., 17 Jul 2025). Subjects S03, S07, S14, and S15 are removed from training to create a validation split, and fine-tuning runs for 60 epochs with batch size 128 (Suzuki et al., 17 Jul 2025).

For 3D training, MotionAGFormer base is used with input sequence length 81 and trained from scratch on AthleticsPose (Suzuki et al., 17 Jul 2025). The paper defines two broad input conditions: GT, using projected 2D ground-truth joints, and DET, using 2D keypoints from the 2D estimator (Suzuki et al., 17 Jul 2025). Within DET, the benchmark distinguishes DET pre-trained, using COCO-pretrained ViTPose, and DET fine-tuned, using ViTPose fine-tuned on AthleticsPose (Suzuki et al., 17 Jul 2025). The 3D model inherits MotionAGFormer default settings but training is extended from 60 to 100 epochs (Suzuki et al., 17 Jul 2025).

The subject split is by athlete rather than by random frames. The test split includes S00 for Discus and Shotput, S05 for Racewalk, S11/S13/S16 for Running, S12/S21/S22/S23 for Starting Dash and Sprint, and S17/S20 for Hurdle; all remaining subjects are used for training (Suzuki et al., 17 Jul 2025). Javelin is excluded from evaluation because it has only one subject (Suzuki et al., 17 Jul 2025).

The benchmark uses standard 2D and 3D metrics alongside downstream kinematic metrics. For 2D pose, the metric is mAP based on Object Keypoint Similarity: $$\text{OKS} = \frac{\sum_{i} \exp(-d_i^2 / 2s^2k_i^2) \cdot \delta(v_i > 0)}{\sum_{i} \delta(v_i > 0)}$$ where $d_i$ is Euclidean keypoint error, $v_i$ is visibility, $s$ is object scale, and $k_i$ is a per-keypoint falloff constant (Suzuki et al., 17 Jul 2025). For 3D pose, the principal metrics are MPJPE and P-MPJPE, defined in prose as average Euclidean joint error and its Procrustes-aligned version (Suzuki et al., 17 Jul 2025).

5. Empirical findings

The central empirical result is that authentic athletics data dramatically improves monocular 3D pose estimation on real athletics motion (Suzuki et al., 17 Jul 2025). On the AthleticsPose test set, a model trained on Human3.6M under DET pre-trained obtains average MPJPE / P-MPJPE of 149.38 / 98.80 mm (Suzuki et al., 17 Jul 2025). Training on AthleticsPose improves this to 41.99 / 31.40 mm under DET pre-trained, and to 40.13 / 30.33 mm under DET fine-tuned (Suzuki et al., 17 Jul 2025). Under GT input, AthleticsPose training yields 29.45 / 21.44 mm (Suzuki et al., 17 Jul 2025).

The comparison to imitated-motion training is even sharper. Under GT input, AthletePose3D yields 115.91 / 88.36 mm, whereas AthleticsPose yields 29.45 / 21.44 mm (Suzuki et al., 17 Jul 2025). This corresponds to an MPJPE reduction of 86.46 mm, about 74.6%, matching the paper’s statement that MPJPE is reduced by approximately 75% (Suzuki et al., 17 Jul 2025).

By action, the DET fine-tuned AthleticsPose model achieves 33.43 / 25.12 mm for Starting Dash, 33.85 / 24.15 mm for Sprint, 30.61 / 22.50 mm for Running, 46.35 / 29.53 mm for Racewalk, 34.43 / 28.14 mm for Hurdle, 63.18 / 49.32 mm for Discus, and 56.73 / 43.66 mm for Shotput (Suzuki et al., 17 Jul 2025). Running is the easiest action, while Discus and Shotput are the hardest; the paper attributes this to the contrast between periodic motions and non-periodic, throwing-intensive actions (Suzuki et al., 17 Jul 2025).

The per-joint analysis shows that errors increase with distance from the root, which is consistent with root-relative lifting (Suzuki et al., 17 Jul 2025). In the DET fine-tuned setting, average MPJPE by major joint is 12.88 mm at Hip, 40.26 mm at Knee, 50.59 mm at Ankle, 38.10 mm at Shoulder, 53.81 mm at Elbow, and 70.17 mm at Wrist (Suzuki et al., 17 Jul 2025). Throwing events are especially problematic in the upper limbs: in Discus, elbow and wrist errors are 91.10 mm and 134.97 mm; in Shotput, 75.80 mm and 100.49 mm (Suzuki et al., 17 Jul 2025).

These findings align with other sports-pose benchmarks that report strong domain dependence. AthletePose3D likewise shows that Human3.6M-trained models generalize poorly to sports motion and that athlete-specific fine-tuning cuts MPJPE from about 214 mm to about 65 mm (Yeung et al., 10 Mar 2025). SportsPose similarly argues that dynamic sports motion requires dedicated datasets with quantitatively validated markerless capture (Ingwersen et al., 2023). AthleticsPose’s distinctive addition is to isolate authentic field-captured athletics as the determining factor (Suzuki et al., 17 Jul 2025).

6. Reliability for sports analysis

AthleticsPose moves beyond standard joint-error reporting by asking whether monocular 3D pose estimates are reliable for sports-relevant kinematics (Suzuki et al., 17 Jul 2025). The paper highlights two major sensitivity factors: camera viewpoint and subject scale. Side-view performance is better than front/back-view performance, with MPJPE / P-MPJPE of 31.42 / 24.12 mm for side view versus 36.34 / 26.53 mm for front/back view (Suzuki et al., 17 Jul 2025). The authors attribute this primarily to subject scale stability: side views preserve more constant image scale, while front/back views involve large depth-induced scale changes (Suzuki et al., 17 Jul 2025).

This interpretation is supported by a scale analysis. Bounding-box height is normalized relative to the maximum observed height, with Large defined as relative scale $^2$0, Medium as $^2$1, and Small as $^2$2 (Suzuki et al., 17 Jul 2025). Pose error worsens monotonically as subjects become smaller: 33.27 / 24.74 mm for Large, 37.15 / 27.57 mm for Medium, and 42.63 / 32.58 mm for Small (Suzuki et al., 17 Jul 2025).

The paper then evaluates two sprint-relevant kinematic indicators. The first is the knee angle of the supporting leg, computed from hip, knee, and ankle coordinates and low-pass filtered (Suzuki et al., 17 Jul 2025). The paper does not print an explicit knee-angle formula, but reports an overall RMSE of $^2$3, with $^2$4 for side view and $^2$5 for front/back (Suzuki et al., 17 Jul 2025). Predicted knee angles retain significant subject-specific patterns, but the method underestimates the magnitude of inter-subject differences and can introduce false distinctions between subjects (Suzuki et al., 17 Jul 2025). This suggests that monocular 3D pose may support coarse subject-specific comparison of posture variables, but not fully trustworthy quantitative ranking without bias correction (Suzuki et al., 17 Jul 2025).

The second case study is maximum knee-drive velocity, derived from the time derivative of inter-knee distance (Suzuki et al., 17 Jul 2025). Here the performance is worse: overall RMSE is 0.77 m/s, with 0.79 m/s for side view and 0.76 m/s for front/back (Suzuki et al., 17 Jul 2025). The estimates show a consistent overestimation bias across subjects, and post-hoc testing indicates failure to reproduce true inter-subject relationships (Suzuki et al., 17 Jul 2025). The authors conclude that current monocular pose error limits practical use for high-speed derivative-based metrics (Suzuki et al., 17 Jul 2025).

This reliability analysis parallels findings in AthletePose3D, which reports stronger correlation for joint-angle waveforms than for velocity waveforms and concludes that velocity estimation remains substantially less reliable than angle estimation (Yeung et al., 10 Mar 2025). A plausible implication is that AthleticsPose should be understood as clarifying a threshold of utility: current monocular methods are conditionally informative for lower-body kinematics in running-related actions, especially under favorable viewpoints and sufficient scale, but are not yet a substitute for high-fidelity biomechanical measurement (Suzuki et al., 17 Jul 2025).

7. Position within the sports-pose literature

AthleticsPose sits at the intersection of several research threads: sports-specific pose datasets, pose-based action/event analysis, and coaching or biomechanics-oriented pose systems. Compared with AthletePose3D, it replaces mixed lab/rink sports capture with authentic athletics-field motion and demonstrates a much larger gap between real and imitated sports training when evaluation is conducted on authentic athletics (Suzuki et al., 17 Jul 2025, Yeung et al., 10 Mar 2025). Compared with SportsPose, it narrows the sports domain to athletics but strengthens the claim of authenticity and practical kinematic relevance (Suzuki et al., 17 Jul 2025, Ingwersen et al., 2023).

The dataset also complements work on pose-based sports-video analysis. Decoupled pose-sequence modeling has been shown effective for event detection in athletics, such as stride-event spotting in long jump and triple jump using pose normalization and temporal convolution (Einfalt et al., 2020). Noisy-pose mining methods have shown that long jump phase segmentation and derived metrics such as run-up duration and step count can be recovered from automatically estimated poses using clustering and HMM decoding (Lienhart et al., 2018). AthleticsPose extends this agenda into monocular 3D evaluation, focusing not on event timing alone but on whether pose estimates are accurate enough for performance variables (Suzuki et al., 17 Jul 2025).

It also intersects with coaching-oriented systems such as PoseCoach, which uses monocular 3D reconstruction, cycle alignment, and customizable kinematic attributes for running analysis (Liu et al., 2022), and with low-data comparison frameworks such as Poze, which use 3D pose, dynamic time warping, and expert-reference statistics for technique feedback (Singh et al., 2024). AthleticsPose differs from these systems by being a dataset-and-benchmark paper rather than an interactive coaching interface or a low-data feedback engine, but its results bear directly on the viability of those applications in athletics (Suzuki et al., 17 Jul 2025).

8. Limitations and open questions

The paper is explicit about several limitations. The participant pool remains modest, with 24 collected and 23 benchmarked subjects, all university student athletes (Suzuki et al., 17 Jul 2025). Event coverage is limited to eight classes, and javelin is underrepresented and excluded from evaluation (Suzuki et al., 17 Jul 2025). Annotation quality depends on markerless motion capture, which, despite consistency checks, can still contain reconstruction errors (Suzuki et al., 17 Jul 2025). The paper does not provide exhaustive calibration equations, per-camera intrinsics, or a formal uncertainty analysis of the markerless ground truth (Suzuki et al., 17 Jul 2025).

Model failure modes are also clear. Performance is view-dependent and scale-dependent, errors grow at distal joints, and upper-limb analysis in throwing events is especially unreliable (Suzuki et al., 17 Jul 2025). Downstream metric estimation exhibits bias: knee-angle estimates underestimate inter-subject differences, while knee-drive velocity is consistently overestimated (Suzuki et al., 17 Jul 2025). This suggests that athletics applications depending on temporal derivatives or detailed upper-limb throwing mechanics remain beyond current monocular reliability (Suzuki et al., 17 Jul 2025).

A common misconception would be to treat strong MPJPE improvements as sufficient evidence of full biomechanical validity. AthleticsPose argues against that interpretation. The paper shows that even when joint-position errors fall substantially, derivative-based metrics and viewpoint robustness remain problematic (Suzuki et al., 17 Jul 2025). Another misconception would be to assume that any “sports dataset” is interchangeable with athletics. The comparison to imitated-motion baselines shows that overlap in nominal action labels does not guarantee transfer if the motion distribution lacks authentic athletic intensity and execution strategy (Suzuki et al., 17 Jul 2025).

The publication of dataset, code, and checkpoints at the project repository is intended to support reproducibility and downstream adoption (Suzuki et al., 17 Jul 2025). A plausible implication is that AthleticsPose may function not only as a benchmark but also as a calibration point for future athletics-specific extensions: more subjects, broader event coverage, richer camera configurations, and evaluations that combine pose accuracy with event timing, contact inference, and technique-specific kinematic validity.