Matteo Campagna
ITEN
Menu
Menu

A pose estimation model can locate an archer's joints from ordinary footage

2 studies · 5 archers measured

ForContextAgainst202420262025, context, n = 42026, context, n = 1
Each circle is a study: its position gives the year, its row whether it is for, against or context, its size how many participants it measured. Light circles are studies on other populations or without archers.

There is only one direct check of the error, from 2026, and it measures little: eight photos of a single archer, one per phase of the shot, with the points found by MediaPipe compared with those marked by hand. The normalised error runs from 0.55 at nocking to 0.71 at anchor, and the authors judge it competitive. But the scale factor is described only through examples, and without it there is no way of telling whether 0.71 is small or large; the declared mean, 0.71, matches the highest value, while the mean of the values in the table is about 0.65; and the photos are still poses chosen to be clear, not a video of a shot with the string and hand close to the face. Before that, in 2025, the coordinates of the shoulders, elbows and wrists extracted from 60-frames-per-second video were enough to tell which of four archers was shooting, a sign that the tracking carries information, but that study does not measure the localisation error. That the model finds the joints can be seen; how well it does so, nobody has yet measured in a way that can be read, which is why the registry keeps the 2026 study as context.

Also phrased as: "pose estimation in archery"; "video analysis with artificial intelligence"; "mediapipe for archers"; "joint detection from video"; "markerless movement analysis"

In practice

One hears that it is enough to film the shot on a phone and run it through a pose recognition program to get the angles of the shoulders, elbows and wrists without needing a laboratory.

Evidence

  • 2025
    Context Archers Published later

    From Full HD video at 60 frames per second of 4 Korean national-level archers, 11 of the 17 keypoints (head, hips, knees, ankles) are excluded because they stay still during the shot, and the remaining 12 coordinates of the shoulders, elbows and wrists are enough to recognise the archer in 98.7 per cent of test cases. The tracking carries information, but the localisation error is not measured.

    Deep Learning Sequence Network for Identifying and Analyzing Archery Shooting Patterns · Jihoon Park, Hyongjun Choi

  • 2026
    Context Archers Published later

    8 reference photos of a single archer, one per phase of the shot, with the points estimated by MediaPipe without retraining compared with those marked by hand: a declared mean normalised error of 0.71, which the authors judge competitive because it falls within the 0.6 to 1.2 range they attribute to earlier studies. The scale factor is not defined exactly, so the number does not tell whether the error is small or large. A regional journal.

    Accuracy Evaluation of 2D MediaPipe-Based Pose Estimation for Archery Posture Detection Using N-MPJPE · Muhammad Andhika Bayu Prasetya, Harits Ar Rosyid, M. Zainal Arifin

  • 2026
    Context Archers Published later

    By phase the error runs from 0.55 at nocking to 0.61 at release, 0.63 when aiming, 0.65 at string hooking and during the draw, 0.68 at set-up, 0.69 in the stance and 0.71 at anchor: it changes little and does not get worse in the more dynamic phases. The mean of these values is about 0.65, while the declared one, 0.71, matches the highest value.

    Accuracy Evaluation of 2D MediaPipe-Based Pose Estimation for Archery Posture Detection Using N-MPJPE · Muhammad Andhika Bayu Prasetya, Harits Ar Rosyid, M. Zainal Arifin

Status history

  1. 2025 Not studied Pose estimation enters archery with a study of four Korean national-level archers: from the coordinates of the shoulders, elbows and wrists extracted from 60-frames-per-second video a network recognises who is shooting in 98.7 per cent of cases, and eleven of the seventeen points are discarded because they stay still during the movement. The tracking carries information, but nobody measures how far the estimated points are from the real joints: on the accuracy of localisation there are still no data.
  2. 2026 Not studied The first check of the error arrives, on eight photos of a single archer in the phases of the shot, with MediaPipe's points compared with those marked by hand: the normalised error runs from 0.55 at nocking to 0.71 at anchor, and the authors judge it competitive. The scale factor, though, is described only through examples, the declared mean of 0.71 is the highest value in the table, whose average is about 0.65, and the images are still poses. One subject and a number that cannot be interpreted do not say how well the model locates the joints, so the study remains context and there is still no usable measurement.

Related

Related articles

Research lines

Where there are no studies on archers, every decision starts from a written hypothesis and is checked against the data over time.

See the full assessment
Cite
APA

Campagna, M. (2026). A pose estimation model can locate an archer's joints from ordinary footage (Version 2). In Evidence Registry. Retrieved 16 October 2026, from https://matteocampagna.com/en/registry/technique/archer-pose-estimation-from-video

BibTeX
@misc{campagna2026_stimaposturadavideo,
  author = {Campagna, Matteo},
  title = {A pose estimation model can locate an archer's joints from ordinary footage},
  year = {2026},
  howpublished = {Evidence Registry, matteocampagna.com},
  url = {https://matteocampagna.com/en/registry/technique/archer-pose-estimation-from-video},
  note = {Version 2; accessed 2026-10-16}
}

Other claims in this domain

See all claims on technique and biomechanics