Pith. sign in

REVIEW 4 major objections 4 minor 13 references

A Real-Time, Vision-Based System for Badminton Smash Speed Estimation on Mobile Devices

T0 review · 4 major / 4 minor · reviewed 2026-08-05 · deepseek-v4-flash

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

Pith's one-line read A smartphone vision system can measure a badminton smash's true peak speed at impact, which consumer radar guns systematically miss.

desk verdict A workmanlike mobile CV pipeline with a plausible story but an internally inconsistent validation: the only ground truth is the same radar gun the paper distrusts. read the letter →

arxiv 2509.05334 v1 pith:VFPAPCYA submitted 2025-09-01 cs.CV cs.MM

Diwen Huang This is my paper · ORCID
classification cs.CVcs.MM
keywords badmintonsmashspeedsmartphonecomputervisionYOLOv5detectionKalmanfilterkinematicestimationradarguncomparisonshuttlecockaerodynamicssportsanalytics
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

This paper claims that a smartphone-based vision system can measure a badminton smash's speed right after impact more accurately than a consumer radar gun. The system detects the shuttlecock in video with a custom-trained object detector, smooths the track with a state-estimation filter, and converts pixel movement to real-world speed using a user-selected court line as a scale. In a 20-smash comparison, the app's peak readings ran systematically higher than the radar gun—about 66 km/h on average—while its readings taken near the net closely matched the radar. The author argues that this pattern is exactly what physics predicts: the shuttlecock decelerates rapidly, so the radar's delayed reading misses the true peak. If the claim holds, it makes professional-grade smash-speed feedback available to any player with a phone.

What carries the argument

The load-bearing mechanism is the spatiotemporal scale factor Sf = d_real/d_pixel, which converts pixel displacement into meters, combined with a tracking pipeline. A custom-trained YOLOv5 detector localizes the shuttlecock frame-by-frame; a linear Kalman filter with a constant-velocity motion model smooths the trajectory; a composite score (0.3 × YOLO confidence + 0.7 × normalized proximity to the Kalman prediction) selects the best detection; and speed is computed from the shuttlecock's leading edge to reduce motion-blur error. The final reported speed is the maximum over consecutive frame pairs.

What would settle it

Set up a high-frame-rate camera (240 fps or more) on the same perpendicular line, record the same smashes, and compute speed over the first few post-impact frames; if the independent peak is materially lower than the app's reported peak, the central claim fails.

Watch

Extended reading notes

Core claim

The paper's central claim is that the vision system's peak-speed readings are not inflated errors but the physically relevant post-impact velocity. Across 20 trials, the app's peak speeds exceeded the radar gun by a mean absolute error of 66.41 km/h (RMSE 74.68 km/h), while the app's estimates of speed near the net, where the radar would be locking on, agreed with the radar. The author interprets the divergence as the shuttlecock's rapid aerodynamic deceleration and the agreement as confirmation that the scale and tracking are calibrated correctly. The conclusion is that the app measures the true peak velocity at impact—something a radar gun cannot do in this setting.

Load-bearing premise

The near-net validation treats the radar gun as ground truth for shuttlecock speed, even though the paper argues radar guns themselves often fail to track shuttlecocks.

Editorial extensions

If this is right

  • For straight smashes, a player or coach can get impact speed with only a phone and a court-line reference—no radar gun, no manual frame-by-frame tracking.
  • Speed comparisons between players or sessions should state whether they refer to post-impact peak or later trajectory speed, because the two can differ by roughly 66 km/h.
  • The near-net agreement supports using the same trajectory to check calibration: if near-net estimates drift from an independent reference, the scale or tracking needs adjustment.
  • The system's demonstration at 30 FPS means it should work on baseline modern phones; higher frame rates are a natural improvement path.

Reading between the lines

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

  • Because the paper's own background section says radar guns often fail to detect shuttlecocks, the near-net agreement could also mean both systems share a late-trajectory bias; an independent high-speed camera reference would settle which reading is right.
  • The 30 FPS sampling means the first tracked frame can occur tens of milliseconds after impact, so even the reported 'peak' is likely a lower bound on the true instantaneous impact speed.
  • The peak-to-net speed difference, combined with published shuttlecock drag measurements, could be used to estimate launch angle or drag coefficient from a single phone video.
  • Relaxing the strict perpendicular-camera assumption with a short tilt calibration would extend the system from straight smashes to cross-court shots.
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 / 4 minor

Summary. The paper presents a smartphone-based computer-vision pipeline for estimating badminton smash speed. A custom YOLOv5 model detects the shuttlecock from a perpendicular camera view; detections are filtered by heuristics, a Kalman filter, and a composite scoring rule; speed is computed from consecutive-frame displacements scaled by a user-selected reference distance. The system is packaged in iOS/Android apps. The authors validate against a Bushnell Speedster III radar gun, reporting that the app's peak speed is systematically higher than the radar and that the app's 'at-net' speed agrees with the radar. They conclude that the system measures the true peak post-impact velocity that radar cannot capture.

Significance. If the validation were sound, this would be a practical and useful contribution: it demonstrates a complete mobile pipeline with a custom detection model (precision 93%, recall 87%, mAP@0.5 91%), explicit handling of perspective scaling, and accessible deployment on consumer devices. The authors are also transparent about the 2D planar assumption and the small validation set. However, the central accuracy claim rests on comparing the system with the same radar instrument that the paper itself argues is unreliable for shuttlecock measurement, and the reported 'at-net' agreement is quantitatively too weak to establish calibration. As it stands, the headline conclusion—that the app measures the true peak velocity—is unsupported.

major comments (4)
  1. [§2.2 and §4.1] The validation is circular. Section 2.2 states that consumer radar guns 'often fail to reliably detect the shuttlecock' due to its non-metallic construction, yet Section 4.1 uses a Bushnell Speedster III radar gun as 'the ground truth' for the at-net comparison. If the radar cannot be trusted for shuttlecocks, agreement with it cannot validate the vision system. The paper needs an independent reference (e.g., high-speed multi-camera triangulation or a calibrated optoelectronic system) or a direct justification of why the radar is reliable specifically near the net.
  2. [Table 1, §4.3] The 'at-net' agreement is not 'remarkable' as claimed. The absolute differences between the app's at-net speeds and the radar range from 0.9 to 34.2 km/h, with MAE ≈ 9.9 km/h and RMSE ≈ 13.2 km/h across 20 trials. Trials 5 and 15 differ by 34.2 and 31.0 km/h, respectively. No repeated trials, standard deviations, or error bars are reported. A measurement that can be off by more than 30 km/h on a 100–150 km/h shot does not establish that the system is 'calibrated correctly.'
  3. [§4.2, Table 1] The comparison is not reproducibly defined. The 'speed at net' is selected by the authors' own system at an 'approximate location,' with no stated criterion for that frame, and no synchronization protocol between the radar gun's lock-on time and the video frame is described. Without a pre-specified operational definition or frame-selection rule, the at-net agreement could reflect post hoc selection. The reader cannot tell whether the comparison is a blind validation or an artifact of the pipeline.
  4. [§3.2.3, Eq. (3)] The 'peak speed' is a frame-to-frame displacement at 30 FPS, i.e., an average velocity over a 33 ms interval, not an instantaneous impact speed. During the first frames after impact the shuttlecock decelerates rapidly, and the reported peak can undershoot the true maximum; motion blur also biases the leading-edge point. The paper provides no evidence that this frame-averaged value equals the true peak. The claim that the system measures 'the true peak velocity at impact' therefore overstates what Eq. (3) computes.
minor comments (4)
  1. [§3.1] The sentence 'demonstrating a high of accuracy' appears to be a typo; it should read 'a high degree of accuracy.'
  2. [References] Reference [9] (Singh et al., 'End-to-end deep learning for directly estimating the Kalman filter gain') is cited in the Kalman-filter paragraph but is not used in the method; either use it or remove it.
  3. [§4.3] The MAE and RMSE for the peak-speed comparison (66.41 and 74.68 km/h) are not meaningful validation metrics if the two systems are explicitly measuring different physical quantities (peak post-impact vs. later radar reading). Report correlation and Bland-Altman agreement for the at-net comparison instead, and include confidence intervals.
  4. [§3.1, §4.1] The custom dataset of 15,000 images is not described in terms of split, annotation protocol, or availability, and the test environment is only a single court with one player. For reproducibility, the authors should provide more detail and ideally release code/data.

Circularity Check

0 steps flagged · score 0.0 of 10

No circular derivation: speed estimates follow from video kinematics, detection, and filtering; the weak radar validation is a correctness concern, not a circularity.

full rationale

The central speed estimate is computed by the kinematic formula vi→i+1 = sqrt((xi+1−xi)^2 + (yi+1−yi)^2) / t_frame × Sf × 3.6 from pixel displacements, a user-supplied scale factor, and frame time (Section 3.2.3). The detection and tracking stages (YOLO, Kalman filter, composite scoring) are independent of the reported speed; no parameter is fitted to the radar gun. The 0.3/0.7 composite weights are set for temporal smoothness, not to reproduce radar readings. The only validation step is the at-net comparison against a Bushnell Speedster III, which the paper itself calls unreliable for shuttlecocks (Section 2.2: "Cheaper consumer models often fail to reliably detect the shuttlecock, as its non-metallic feather or nylon construction provides a poor reflective surface for radar waves"). That is a serious validity problem for the claim "This confirms that our system is calibrated correctly," but it is not circular: the at-net speed is produced by the same pipeline and compared, not defined as equal to the radar reading. The limitations section (5.1) concedes the small sample and sub-maximal smashes, which affects statistical support rather than self-justification. No self-citation, imported uniqueness, or renamed result is load-bearing. Therefore no circular step can be exhibited.

Assumptions & free parameters 6 free parameters · 5 assumptions · 0 invented entities

The system's output depends on a user-provided scale factor, several hand-set detection thresholds, and the unstated Kalman noise covariances; none are fitted to an independent reference, and the validation reference (radar) is itself questioned by the paper.

free parameters (6)
  • YOLO confidence threshold = 0.1
    Chosen to maximize recall; low threshold admits more false positives.
  • IoU threshold for NMS = 0.45
    Non-max suppression threshold in YOLO; hand-selected.
  • Heuristic speed bounds = 5 and 375 km/h
    Detections implying speeds outside this range are discarded as implausible; bounds are arbitrary.
  • Composite scoring weights = 0.3 (YOLO conf), 0.7 (Kalman prox)
    Weights 'set empirically' to prioritize temporal smoothness; no validation.
  • Proximity normalization factor = 0.25 * frame width
    Hand-set denominator for Kalman proximity score; affects detection selection under motion.
  • Kalman filter noise covariances
    Not reported; the Q and R matrices affect smoothing strength and are unstated.
assumptions (5)
  • domain assumption The camera is positioned so the shuttlecock moves entirely in a plane parallel to the image plane (no z-axis motion).
    Section 3.1 and 3.2.1 require a perpendicular view; any depth motion underestimates speed. Acknowledged as a limitation but central to the method.
  • domain assumption The user-selected court line provides a correct pixel-to-meter scale for the shuttlecock's trajectory.
    Section 3.2.1: user picks two points on a line at the same depth as impact; user error directly scales every speed.
  • standard math The Kalman filter's constant-velocity model adequately represents shuttlecock motion between frames.
    Section 3.2.2: standard tracking assumption; shuttlecock decelerates significantly, so constant velocity is a rough approximation at 30 FPS.
  • domain assumption The radar gun is an accurate reference for speed near the net.
    Section 4.1 uses radar as 'ground truth'; contradicts Section 2.2's claim that radar is unreliable for shuttlecocks.
  • domain assumption The motion-compensated leading edge point reliably locates the shuttlecock under motion blur.
    Section 3.2.3; no explicit validation of this feature point's accuracy.

how reviews work

0 comments
Cite this review

Pith. "Pith review of A Real-Time, Vision-Based System for Badminton Smash Speed Estimation on Mobile Devices." pith.science (2026). https://pith.science/paper/VFPAPCYA

@misc{pith2026250905334,
  author       = {Pith},
  title        = {Pith review of: A Real-Time, Vision-Based System for Badminton Smash Speed Estimation on Mobile Devices},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/VFPAPCYA}},
  note         = {Machine review of arXiv:2509.05334}
}
read the original abstract

Performance metrics in sports, such as shot speed and angle, provide crucial feedback for athlete development. However, the technology to capture these metrics has historically been expensive, complex, and largely inaccessible to amateur and recreational players. This paper addresses this gap in the context of badminton, one of the world's most popular sports, by introducing a novel, cost-effective, and user-friendly system for measuring smash speed using ubiquitous smartphone technology. Our approach leverages a custom-trained YOLOv5 model for shuttlecock detection, combined with a Kalman filter for robust trajectory tracking. By implementing a video-based kinematic speed estimation method with spatiotemporal scaling, the system automatically calculates the shuttlecock's velocity from a standard video recording. The entire process is packaged into an intuitive mobile application, democratizing access to high-level performance analytics and empowering players at all levels to analyze and improve their game.

Figures

Figures reproduced from arXiv: 2509.05334 by the authors.

Figure 1
Figure 1. The end-to-end system pipeline. The user uploads a video, which is processed frame-by-frame by the [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Comparison of camera views. (a) The required perpendicular view for our 2D analysis. (b) A typical broadcast view, unsuitable due to perspective distortion. false positive. Similarly, any detection implying a speed over 375 km/h is rejected to favor precision over recall for extreme outliers. Kalman Filter Smoothing. To incorporate tem￾poral context, we employ a linear Kalman filter [8, 9], which assumes a constant … view at source ↗
Figure 3
Figure 3. Scatter plot of speeds measured by our system versus the radar gun. The consistent position of points above the y=x line demonstrates the systematic difference in measurement, with our system capturing the higher peak velocity at impact. the moment of impact, our system avoids the significant underestimation caused by the shuttlecock’s rapid de￾celeration, a factor that fundamentally limits the utility of radar-base… view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

13 extracted references · 13 canonical work pages

  1. [1]

    Badminton second to soc- cer in participation worldwide,

    B. Clement, “Badminton second to soc- cer in participation worldwide,” ESPN.com, Jul. 2004, accessed: 2025-08-31. [Online]. Available: https://www.espn.com/olympics/ summer04/badminton/news/story?id=1845228

    work page 2004

  2. [2]

    Limitations of doppler radar for small, low-velocity projectiles,

    V. Edrich, “Limitations of doppler radar for small, low-velocity projectiles,” Sports Engineer- ing, vol. 15, no. 2, pp. 89–95, 2012

    work page 2012

  3. [3]

    Badminton shut- tlecock tracking and 3d trajectory estimation from video,

    C.-L. Lee and P. J. Ramadge, “Badminton shut- tlecock tracking and 3d trajectory estimation from video,” Princeton University, Tech. Rep., 2019, princeton DataSpace, Electrical Engineering

    work page 2019

  4. [4]

    Reconstruc- tion of 3d ball/shuttle position by two image points from a single view,

    L. Shen, Q. Liu, L. Li, and Y. Ren, “Reconstruc- tion of 3d ball/shuttle position by two image points from a single view,” in Proceedings of the 11th In- ternational Symposium on Computer Science in Sport (IACSS 2017) , ser. Advances in Intelligent Systems and Computing, M. Lames, D. Saupe, and J. Wiemeyer, Eds., vol. 663. Springer Interna- tional Publish...

    work page 2017

  5. [5]

    Track- netv2: Efficient shuttlecock tracking network,

    N.-E. Sun, Y.-C. Lin, S.-P. Chuang, T.-H. Hsu, D.-R. Yu, H.-Y. Chung, and T.-U. Ik, “Track- netv2: Efficient shuttlecock tracking network,” in Proceedings of the 2020 International Conference on Pervasive Artificial Intelligence (ICPAI) , Dec. 2020, pp. 86–91

    work page 2020

  6. [6]

    Jocher, “YOLOv5,” GitHub repository,

    G. Jocher, “YOLOv5,” GitHub repository,

    work page

  7. [7]

    You Only Look Once: Unified, real-time object detection,

    J. Redmon, S. Divvala, R. Girshick, and A. Farhadi, “You Only Look Once: Unified, real-time object detection,” in Proceedings of the IEEE Conference on Computer Vision and Pat- tern Recognition (CVPR), 2016, pp. 779–788

    work page 2016

  8. [8]

    A new approach to linear filtering and prediction problems,

    R. E. Kalman, “A new approach to linear filtering and prediction problems,” Journal of Basic Engi- neering, vol. 82, no. 1, pp. 35–45, 1960. 5

    work page 1960

Show all 13 references
  1. [10]

    Simple online and realtime tracking,

    A. Bewley, Z. Ge, L. Ott, F. Ramos, and B. Upcroft, “Simple online and realtime tracking,” arXiv preprint arXiv:1602.00763, 2016. [Online]. Available: https://arxiv.org/abs/1602.00763

    2016 arXiv

  2. [11]

    The physics of badminton,

    C. Cohen, B. D. Texier, D. Qu´ er´ e, and C. Clanet, “The physics of badminton,” New Journal of Physics, vol. 17, no. 6, p. 063001, 2015

    2015

  3. [12]

    Monotrack: Shut- tle trajectory reconstruction from monoc- ular badminton video,

    P. Liu and J.-H. Wang, “Monotrack: Shut- tle trajectory reconstruction from monoc- ular badminton video,” arXiv preprint arXiv:2204.01899, 2022. [Online]. Available: https://arxiv.org/abs/2204.01899 6

    2022 arXiv

  4. [2018]

    Available: https://arxiv.org/abs/ 1802.05389

    [Online]. Available: https://arxiv.org/abs/ 1802.05389

    arXiv

  5. [2020]

    Available: https://github.com/ ultralytics/yolov5

    [Online]. Available: https://github.com/ ultralytics/yolov5

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.