Pith. sign in

REVIEW 4 major objections 5 minor 11 references

Markerless Stride Length estimation in Athletic using Pose Estimation with monocular vision

T0 review · 4 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 A homography built from track-line intersections and pose-detected toe points yields markerless stride-length estimates from monocular video.

desk verdict A practical-looking stride-length pipeline undone by an unstated homography scale; the consistency check is real but relative, so the absolute metric claim is unverified. read the letter →

arxiv 2507.03016 v1 pith:FUMJOXEZ submitted 2025-07-02 cs.CV

Patryk Skorupski , Cosimo Distante , Pier Luigi Mazzeo This is my paper
classification cs.CV
keywords poseestimationstridelengthhomographymonocularvisiontracklinedetectionmarkerlessmotionanalysisathleticsvideoanalyticsprobabilistichoughtransform
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 aims to show that a runner's stride length can be measured from ordinary monocular video without markers, using pose estimation to find the feet and a homography built from track-line intersections to convert image positions into metric distances. The authors combine edge detection and a probabilistic Hough transform to find the track's horizontal and vertical lines, use their intersections to define a homography, and then project toe keypoints (detected by a pose model and refined by a stationarity rule) into a top-down world plane. Across three athletes, each filmed in two separate videos, the average stride lengths are stable enough (about 1.66–1.90 m, with 2–6 percent variation between videos) that the authors argue the tool is usable for coaching and training. If true, this gives coaching staff a low-cost, unobtrusive way to monitor gait parameters like stride length and pace transition in practice and competition.

What carries the argument

The load-bearing mechanism is the planar homography matrix $H \in \mathbb{R}^{3\times 3}$ that maps image-plane pixel coordinates to world-plane metric coordinates using at least four point correspondences. Here the correspondences come from intersections between horizontal and vertical track lines detected by a probabilistic Hough transform, with the homography recomputed per frame and the median matrix used to suppress outliers. A second component is the contact-point detector: a toe keypoint from the EfficientPose pose model is treated as a foot landing when it stays in the same pixel for two consecutive frames, and each detected point is projected along the line through the vanishing point before being mapped by the homography.

What would settle it

Film the same athlete on a track while a motion-capture system or force plates record true stride length; if the homography-based estimates differ from the ground truth by more than the reported run-to-run variability, the metric scale or the contact detection is wrong. A simpler check is to place two markers a known distance apart on the track inside the camera view and compare the homography-reconstructed distance to that known value.

Watch

Extended reading notes

Core claim

The central claim is that a planar homography estimated from the intersections of detected track lines turns a monocular camera view into a metric top-down view, and that Euclidean distances between consecutive foot-landing points in that view are stride lengths. The authors report that the method yields reproducible per-athlete estimates across two videos—1.66 vs 1.76 m for one runner, 1.87 vs 1.80 m for a second, and 1.90 vs 1.86 m for a third—and take this consistency, in the absence of ground truth, as evidence of reliability. The intended conclusion is that markerless video analysis can support training and competition monitoring of athletes.

Load-bearing premise

The method assumes the real-world spacing and position of the track lines used to build the homography are known and correct, because that is the only source of metric scale; if the assumed line geometry is wrong, every stride estimate is scaled by the same wrong factor and the cross-video consistency cannot reveal the error.

Editorial extensions

If this is right

  • Coaches can obtain per-athlete stride-length curves from a single fixed camera, without markers or wearables, as long as track lines are visible.
  • Because the metric scale sits in the homography, the same pipeline works for any track whose line spacing is known; no per-athlete calibration is required.
  • The combination of classical line detection with a real-time pose model points to a practical pipeline that can run near video frame rate.
  • If stride length is stable across videos for a given athlete, then changes in the estimate can flag fatigue, pacing adjustments, or speed changes for a coach.

Reading between the lines

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

  • The method's absolute accuracy is untested: without ground truth, a constant scale error from wrong assumed line positions would hide in the cross-video consistency; a validation against motion capture or known-distance markers would settle the metric scale.
  • The two-frame stationarity rule fixes an implicit temporal resolution; at 30 fps, very brief ground contacts in sprinting may be missed, so high-frame-rate capture is a natural stress test.
  • Because the homography is derived from track lines in each frame, camera motion or zoom that keeps lines visible could in principle be tolerated, extending the method beyond fixed cameras.
  • The paper's future step of adapting this to non-track settings is the logical next move: replace the line-based homography with SLAM or depth information, and the rest of the pose-to-stride pipeline carries over.
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 proposes a markerless pipeline for estimating stride length from monocular video of runners on an athletics track. The method combines Canny edge detection, probabilistic Hough transforms, vanishing-point estimation, and a planar homography to map image points to metric coordinates, then uses EfficientPose keypoints with a stationarity heuristic to identify foot contacts and computes Euclidean distances in the transformed plane. Experiments on three athletes, with two videos each, report average stride lengths of 1.66–1.90 m with cross-video differences of 0.04–0.10 m; no ground truth is available. The conclusion states that the results confirm the method's reliability.

Significance. If validated, the approach would provide a low-cost, markerless tool for stride monitoring in athletics, and the combination of classical geometric line detection with a neural pose estimator is a reasonable engineering choice. The paper is transparent about the absence of ground truth and lists validation as future work, which is a strength. However, the current evidence supports only consistency under a fixed camera/track setup, not accuracy or reliability in metric units; the unstated world-plane calibration is a critical gap. The contribution is therefore at the feasibility-demonstration level rather than a validated measurement method.

major comments (4)
  1. [Section 2.4, Eq. (1)] The homography in Section 2.4 is computed from point correspondences between pixel coordinates and “metric coordinates” in R^2, but the manuscript nowhere states the actual real-world coordinates or measured distances assigned to the four track intersections (derived from the fourth and fifth horizontal lines in Section 2.3.1). Since H defines the absolute scale of every stride length, an error of factor k in the assumed track geometry multiplies all reported distances by k. The cross-video consistency in Table 1 is invariant under this global scaling, so it cannot validate the absolute values 1.66–1.90 m. Please specify the measured world coordinates and how they were obtained; otherwise the method is not reproducible and the metric claim is unverified.
  2. [Section 3, Table 1] The only quantitative evidence is consistency of average stride lengths between two videos for three athletes, with differences of 0.04–0.10 m. The text itself states “Since there is no ground truth … we can only compare the consistency between the two videos,” and the statement that “the foot position is ever correctly detected for each steps” (presumably “never correctly detected for every step”) indicates unquantified detection errors. Without ground-truth measurements or any error bar, consistency does not establish accuracy; it is fully compatible with a constant systematic error. The concluding sentence of Section 4 (“The results obtained confirm the reliability of the method”) is therefore not supported by the reported experiments. Add a ground-truth comparison (e.g., motion capture, manual stride marking, or known track distances) or substantially weaken the claim to feasibility.
  3. [Section 2.5] The rule that a toe keypoint remaining at the same pixel for two consecutive frames is a foot strike is asserted without tolerance, validation, or sensitivity analysis. Pose-estimation jitter, the 30 fps frame rate, and camera perspective all affect this test, and mis-detected contacts directly change the Euclidean distances in the transformed plane. Please report the distribution of detected contact points, compare against manual annotations, or analyze sensitivity to the stationarity threshold.
  4. [Section 3 (homography aggregation)] The paper states that “the median of the matrices computed for each individual frame was taken” to obtain the final homography, but it does not define the median operation. Homography matrices form a projective group rather than a vector space, so an element-wise median is not geometrically meaningful and may yield a matrix that does not represent a valid plane transformation. Please define the aggregation procedure and show that the reported stride lengths are stable under alternative choices.
minor comments (5)
  1. [Section 2.3.2] The sentence “Notice that if ‘inclinate’ was intended to mean ‘near-vertical’ …” is an unedited editorial or translation note and should be removed from the final text.
  2. [Section 2.3.1] The reference to “Algorithm ??” is incomplete; either provide the pseudocode or remove the reference.
  3. [Affiliation] The affiliation contains a typo: “sIstemi” should be “Sistemi”.
  4. [Section 3] The phrase “the foot position is ever correctly detected for each steps” is grammatically incorrect and ambiguous; it should be revised to “the foot position is not always correctly detected for every step”.
  5. [Figure 12 and Table 1] Figure 12 is described as showing results for two runners, while Table 1 reports three athletes; clarify whether the third athlete's result is omitted from the figure.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the stride-length values are computed from pose keypoints via a homography built from track-line correspondences, with no parameter fitted to stride data and no load-bearing self-citation.

full rationale

The derivation chain is self-contained as a measurement pipeline, not a circular argument. Section 2.3 detects horizontal and vertical track lines using Canny edge detection and Probabilistic Hough Transform; Section 2.3.1 selects four reference points from intersections of the fourth and fifth horizontal lines with vertical lines; Section 2.4 uses these correspondences to compute a planar homography (Eq. 1) mapping image pixels to metric world coordinates. Section 2.5 obtains foot landing points from EfficientPose keypoints plus a stationarity criterion, and Section 3 projects these points and computes Euclidean distances in the transformed plane to obtain stride lengths. Nothing in this chain is fitted to the reported stride values: the homography is fit to track geometry, not to athlete stride measurements, and the stride lengths are computed from the transformed keypoints rather than being recovered by optimizing any parameter against them. The cross-video consistency check is a weak validation because the same geometric calibration could carry a systematic scale error, but that is a reproducibility or ground-truth gap, not circularity. There are no load-bearing self-citations: the cited pose estimators and prior systems are external tools, not author-specific results invoked to force the conclusion. An unusual editorial note in Section 2.3.2 ('Notice that if "inclinate" was intended...') and a missing Algorithm cross-reference ('Algorithm ??') are manuscript-quality issues; they do not indicate that the derivation reduces to its inputs. Accordingly, no circular step meets the evidentiary bar.

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

The method rests on hand-set Hough thresholds, angle tolerances, and grouping thresholds that determine the homography; the metric world coordinates of the track lines are assumed known but never specified; and the foot-contact heuristic is an unvalidated assumption. No new physical entities are introduced.

free parameters (5)
  • Hough transform thresholds (minLineLength, maxLineGap) = not reported
    Set by hand in Sections 2.2 and 2.3.1; these control which segments are classified as track lines and therefore determine the homography.
  • Horizontal line angle tolerance = ±10°
    Section 2.3.1: used to classify lines as horizontal before grouping and merging.
  • Vertical line angle threshold and grouping threshold = 2° and 15°
    Section 2.3.2: the 2° threshold identifies near-vertical lines, and the 15° threshold groups them to select the dominant line.
  • Adaptive grouping thresholds per distance area = not specified
    Section 2.3.1: thresholds are reduced in distant image regions, but the actual values are not reported.
  • Stationarity window for foot contact = 2 frames
    Section 2.5: a toe keypoint must remain in the same pixel for two consecutive frames to be considered a landing point.
assumptions (5)
  • standard math A planar homography can be estimated from four point correspondences between the image plane and a world plane.
    Section 2.4: the projective geometry assumption underlying the metric reconstruction.
  • domain assumption The track surface inside the region of interest is approximately planar and the camera follows a pinhole model with negligible lens distortion.
    Sections 2 and 3: the homography maps the image to a metric plane, but no distortion correction or non-planarity handling is mentioned.
  • domain assumption The real-world metric coordinates of the track lines are known.
    Section 3: the homography requires metric correspondences, but the paper never states where these measurements come from.
  • domain assumption EfficientPose's toe keypoints correspond to the anatomical toe and are tracked reliably in running scenes.
    Section 2.5: foot positions are based on toe landmarks, but no accuracy evaluation of the pose model is provided.
  • ad hoc to paper A toe keypoint that stays in the same pixel for two consecutive frames indicates a foot strike.
    Section 2.5: this heuristic is used to select landing points and is not validated against force plates or motion capture.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Markerless Stride Length estimation in Athletic using Pose Estimation with monocular vision." pith.science (2026). https://pith.science/paper/FUMJOXEZ

@misc{pith2026250703016,
  author       = {Pith},
  title        = {Pith review of: Markerless Stride Length estimation in Athletic using Pose Estimation with monocular vision},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/FUMJOXEZ}},
  note         = {Machine review of arXiv:2507.03016}
}
read the original abstract

Performance measures such as stride length in athletics and the pace of runners can be estimated using different tricks such as measuring the number of steps divided by the running length or helping with markers printed on the track. Monitoring individual performance is essential for supporting staff coaches in establishing a proper training schedule for each athlete. The aim of this paper is to investigate a computer vision-based approach for estimating stride length and speed transition from video sequences and assessing video analysis processing among athletes. Using some well-known image processing methodologies such as probabilistic hough transform combined with a human pose detection algorithm, we estimate the leg joint position of runners. In this way, applying a homography transformation, we can estimate the runner stride length. Experiments on various race videos with three different runners demonstrated that the proposed system represents a useful tool for coaching and training. This suggests its potential value in measuring and monitoring the gait parameters of athletes.

Figures

Figures reproduced from arXiv: 2507.03016 by the authors.

Figure 1
Figure 1. Pipeline of the proposed methodology to produce thorough and measurable data.5, 6 This oversight highlights the need for more reliable evaluation techniques that can provide accurate and useful information about athletes’ overall performance and technical execution. Deep learning models and machine vision technology have opened up new avenues for improving evaluation efficiency and accuracy. A model based on pose es… view at source ↗
Figure 2
Figure 2. Training area which contains the lines [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Edge extracted from the track. 2 Materials and Method Our method can be divided into different steps as schematized in figure 1. Edge Detection: Analyzing the image to detect sharp changes in brightness (edges), which helps identify the boundaries of the track or lane lines. In this way, we can isolate the most prominent lines in the scene for further analysis. Probabilistic Hough Transform: A voting algorithm that … view at source ↗
Figures from the paper (9 more)
Figure 5
Figure 5. Figure 5: figure 5 [PITH_FULL_IMAGE:figures/full_fig_p004_5.png]
Figure 4
Figure 4. Figure 4: Segments grouped by area, the reference segment in red, the thresholds for each of the three areas in white [PITH_FULL_IMAGE:figures/full_fig_p004_4.png]
Figure 6
Figure 6. Figure 6: Vertical and Horizontal lines detected by the proposed method. [PITH_FULL_IMAGE:figures/full_fig_p005_6.png]
Figure 5
Figure 5. Figure 5: Merged segments grouped by area [PITH_FULL_IMAGE:figures/full_fig_p005_5.png]
Figure 7
Figure 7. Figure 7: Vanishing point detection using identified vertical lines. [PITH_FULL_IMAGE:figures/full_fig_p006_7.png]
Figure 8
Figure 8. Figure 8: The white circle identified the detected landmarks for Athlete pose estimation. The 33 landmarks represent [PITH_FULL_IMAGE:figures/full_fig_p007_8.png]
Figure 9
Figure 9. Figure 9: In yellow all the key points detected while in red those considered for the distance calculation. [PITH_FULL_IMAGE:figures/full_fig_p007_9.png]
Figure 10
Figure 10. Figure 10: First Runner [PITH_FULL_IMAGE:figures/full_fig_p008_10.png]
Figure 11
Figure 11. Figure 11: Second Runner [PITH_FULL_IMAGE:figures/full_fig_p008_11.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

11 extracted references · 11 canonical work pages

  1. [1]

    Effect analysis of functional physical fitness training based on improved genetic algorithm under functional analysis

    Danxu Lu. Effect analysis of functional physical fitness training based on improved genetic algorithm under functional analysis. Molecular & Cellular Biomechanics, 22(2):1044, Jan. 2025

    work page 2025

  2. [2]

    Influence of nutritional supplementation and sports training on the physical fitness of track and field athletes

    Wenbing Zhu. Influence of nutritional supplementation and sports training on the physical fitness of track and field athletes. Molecular & Cellular Biomechanics, 19(2):89–96, 2022

    work page 2022

  3. [3]

    Study on the influence of training methods on track and field athletes’ performance

    Duan Mingtaou. Study on the influence of training methods on track and field athletes’ performance. Journal of Sport Psychology/Revista de Psicología del Deporte, 32(2), 2023

    work page 2023

  4. [4]

    An improved logistic regression method for assessing the performance of track and field sports

    Songling Zheng and Xi Man. An improved logistic regression method for assessing the performance of track and field sports. Computational Intelligence and Neuroscience, 2022(1):6341495, 2022

    work page 2022

  5. [5]

    Human skeleton key point detection method based on openpose-slim model.Journal of Computer Applications, 39(12):3503, 2019

    J Wang and J Li. Human skeleton key point detection method based on openpose-slim model.Journal of Computer Applications, 39(12):3503, 2019

    work page 2019

  6. [6]

    Feasibility of openpose markerless motion analysis in a real athletics competition

    Neil J Cronin, Josh Walker, Catherine B Tucker, Gareth Nicholson, Mark Cooke, Stéphane Merlino, and Athanas- sios Bissas. Feasibility of openpose markerless motion analysis in a real athletics competition. Frontiers in Sports and Active Living, 5:1298003, 2024

    work page 2024

  7. [7]

    Deep learning models for digital image processing: a review

    R Archana and PS Eliahim Jeevaraj. Deep learning models for digital image processing: a review. Artificial Intelligence Review, 57(1):11, 2024

    work page 2024

  8. [8]

    Detect2Interact: Localizing Object Key Field in Visual Question Answering with LLMs

    Jialou Wang, Manli Zhu, Yulei Li, Honglei Li, Longzhi Yang, and Wai Lok Woo. Detect2Interact: Localizing Object Key Field in Visual Question Answering with LLMs . IEEE Intelligent Systems, 39(03):35–44, May 2024

    work page 2024

Show all 11 references
  1. [9]

    Efficientpose: Efficient human pose estimation with neural architecture search

    Wenqiang Zhang, Jiemin Fang, Xinggang Wang, and Wenyu Liu. Efficientpose: Efficient human pose estimation with neural architecture search. Computational Visual Media, 7:335–347, 2021

    2021

  2. [10]

    Z. Cao, G. Hidalgo Martinez, T. Simon, S. Wei, and Y . A. Sheikh. Openpose: Realtime multi-person 2d pose estimation using part affinity fields. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2019

    2019

  3. [11]

    Movenet: A deep neural network for joint profile prediction across variable walking speeds and slopes

    Rishabh Bajpai and Deepak Joshi. Movenet: A deep neural network for joint profile prediction across variable walking speeds and slopes. IEEE Transactions on Instrumentation and Measurement, 70:1–11, 2021. 9

    2021

Pith tools

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