Machine learning and computer vision in archery research

What a learning program does

Machine learning is not a measuring instrument: it works on data from cameras and sensors and learns a task from examples that have already been solved. In archery research the tasks are of two kinds. The first is taking over a slow manual measurement: finding the joints in a video, timing the phases, reading the holes in the target face or marking the quiet eye. The second is to classify or predict: who is shooting, who has more potential, how many points someone will score. How measurement left the laboratory is told in the section Measuring the shot outside the lab further down; this page looks at how an accuracy figure should be read.

Finding the body in a video

Pose estimation programs such as MediaPipe, YOLOv8 or RTMO look for 17 to 33 body points in each frame. They are trained on people in general, not archers. The only direct check of their error concerns eight photos of a single archer: the normalised error runs from 0.55 at nocking to 0.71 at anchor, but the scale factor is not defined, and without it there is no telling whether that is small or large. This is why the entry on pose estimation from video stays at "not studied".

The coordinates are also two-dimensional, and a side view like the one a 2025 study used to measure the angles of 104 students at the release cannot see rotations. Frame rate matters as well: at 30 frames per second the release, which takes about fifteen thousandths of a second from its start to the arrow leaving the string, falls entirely between two images. The strongest result therefore concerns a long interval: the time from anchor to release, measured on competition footage with a mean error of 0.34 s against an expert, provided the start of the anchor is detected from the angle of the drawing arm; with the string near the face alone, the error rises to 1.4 s. This is the entry on automatic recognition of the shot phases, at "reported".

Targets and signals

On the target, the most accurate system locates the holes in a 40 cm face with a mean error of 1.41 mm, but on photos from a single archer and a single range; the entry on scoring from images is "contested", because the only field test compared with the judges detects 91 arrows out of 100 and scores 67 of them correctly. The other systems are on the page on automatic target scoring.

With sensors the program sees a curve rather than the body: a gyroscope on the bow of five recurve archers was used to group similar shots and to show at which instant two shots part company, in windows of about half a second. Electro-oculography is a signal too: on the quiet eye of ten club archers the two best networks reach about 93 and 92%, and the best drops to 84% on the raw trace (automatic quiet eye measurement).

Where accuracy gets inflated

An accuracy figure states how often the model got it right on data it had not seen, and the first check is exactly that. Frames of the same shot, or shots from the same session, are nearly identical, and if they end up on both sides the model recognises them instead of generalising. In the Indonesian study on the phases the random forest reached 100% on the training data, and the paper does not say that the archers were kept apart; the same doubt applies to the four Korean archers and to the wristband, which goes from 96.0% on windows of signal to 75% on single samples, the measure the authors put forward for real use.

The clearest case is the Indian study of 32 archers that predicts the score of a 30-arrow session. The same archers sit in both training and test sets, and seven of the ten most important variables are fixed characteristics of theirs, the score at the previous competition among them. The declared accuracy is 98.18%, but in the same table a model with an R² of 0.000, which explains nothing, reaches 97.27%: with scores all close to 279 out of 300, guessing the mean is enough to land close. The figure to read is the best model's R² of 0.575, and the entry on how scores are predicted is "contested".

The second check concerns the labels. In the Malaysian studies of young archers the high and low potential groups were built from the same physical measures and the same score that the classifiers then use to find them again: the 97.5% reached by support vector machines and the 82.5% of a weighted k-NN measure the consistency of that procedure, not talent, as the entry on physical tests for talent selection says.

The third is the reference: if the measurement to be imitated is an expert's judgement, the network learns that expert's imprecision too. The last is the size of the test set: in the student study, judging by the per-class figures, the 80% rests on ten people, and the validation estimate of 72.3% is more cautious.

What it can tell an archer or a coach

These programs currently do two things well. They time the interval from anchor to release from competition footage with an error of about a third of a second, enough to count competition timings but not to judge an arrow. And they show that each archer's shot has a shape of its own: from the trajectories of shoulders, elbows and wrists alone a network identifies which of four archers is shooting in 98.7% of test cases, although it may partly rely on build and framing (identifying the archer from movement). This supports comparing each shot with the same archer's usual pattern, in line with the evidence that took the entry on one correct elite technique to "refuted".

They cannot yet say whether a departure from the usual pattern makes the shot worse. The Korean study flags the joints furthest from that archer's average, but none of its measurements is linked to the score or to where the arrow landed, and on scores nobody has yet compared a model with the simplest prediction, the average of recent competitions.

What to check

• which instrument produced the data, at how many frames or samples per second, and whether the measurement is in two or three dimensions;
• what the program is compared with: an independent measurement, one or more experts, no reference at all;
• whether the archers, sessions and shots used for training are kept apart from those used for testing;
• where the labels come from, and whether they contain the variables the model uses;
• how many cases the test set holds, and what a model that always predicts the mean would score;
• whether any figure is linked to the score or to where the arrow landed.

How design and sample are weighed is explained in How we read a study, the timing of the shot in how long a shot lasts; for archers who film themselves on a phone there is the article on video analysis and for selecting young archers the one on talent identification.

@misc{campagna2026_enresearchmethodsmachinelearningcomputer,
  author = {Campagna, Matteo},
  title = {Machine learning and computer vision in archery research},
  year = {2026},
  howpublished = {Research, matteocampagna.com},
  url = {https://matteocampagna.com/en/research/methods/machine-learning-computer-vision},
  note = {accessed 2026-10-16}
}