MediaPipe and archer posture: Prasetya 2026 study

The study checks how well MediaPipe, free software that picks out the joints in an image, locates an archer's body in the eight phases of the shot the authors distinguish: stance, nocking, hooking, set-up, drawing, anchoring, aiming and release. Each phase has a single photograph of the same athlete, and the points the software finds are compared with points marked by hand on the same picture.

Error is expressed as N-MPJPE, the mean distance between estimated and reference points divided by a body measurement. By phase it runs from 0.55 for nocking to 0.71 for anchoring; the other values are 0.69 for stance, 0.65 for both hooking and drawing, 0.68 for set-up, 0.63 for aiming and 0.61 for release. The authors report a mean of 0.71 and call it competitive, since it falls inside the 0.6 to 1.2 range they attribute to earlier studies. An experienced coach then looked at the output and found it consistent with the technique as taught.

MediaPipe is used as distributed, already trained, with no tuning on archery images: it returns 33 body points, including shoulders, elbows, wrists, hips and knees. The eight photographs were taken indoors under controlled lighting, with the camera about 3 to 5 metres from the athlete at body height; blurred images, or ones with parts of the body hidden, were discarded. Phases were labelled by hand, and a coach confirmed that each photograph showed a correct execution of its phase.

The error for each joint is the pixel distance between the estimated and the annotated point, divided by a scale factor taken from the reference pose; the paper describes it by example (the shoulder-to-hip distance, or the largest distance between two joints) without saying which was used. Averaging over all joints gives the value for the phase. The work follows a six-step framework borrowed from IT system development (prepare, plan, design, implement, operate, optimise), and in the last step some operating parameters were adjusted before the final evaluation.

The sample is eight images of one archer, whose level, discipline, sex and shooting side are not given. The paper does not say who marked the reference points or how many times, so the precision of the reference itself is unknown. The scale factor is not defined exactly, and without it the figure of 0.71 cannot be compared with other studies; the 0.6 to 1.2 range attributed to the literature comes without specific references either. If the factor were the shoulder-to-hip distance, a mean error of 71% of that distance would be very large for a still image: the number alone does not say whether the error is small or large.

The figures also fail to agree with each other. The mean of the phase values in the table is about 0.65, while the stated mean, 0.71, matches the highest value, the one for anchoring. The text explains that static phases give smaller errors and dynamic ones larger, yet stance has the second highest error (0.69) and release one of the lowest (0.61). The photographs are still poses chosen for clarity: they say nothing about how the software copes with actual shooting footage, with the string, the bow arm and the hand close to the face.

For anyone who films themselves on a phone and uses an app that draws a skeleton over the body, this paper does not yet offer a measurement to rely on. It shows that MediaPipe finds the main joints in still, well-lit photographs of an archer and that the error changes little from phase to phase; it does not say by how many degrees an elbow or shoulder angle might be off, which is what a coach wants to know. The differences that matter in shooting, such as a drawing elbow a few centimetres higher or lower at anchor, may be smaller than the software's error, and the study gives no way of telling. Using these tools to correct posture would need tests on more archers, on video, against an independently measured reference.