Anchor to release timed by AI: Lee 2024 study

The authors present a programme that measures by itself, from video, how long passes between anchor and release. They tested it on 58 shots taken from public broadcasts on YouTube (the Paris 2024 Olympics and a special match of the Korean national team), comparing it with manual timing by a video analysis expert using Dartfish software.

The start of the anchor was located with a mean error of about 0.29 seconds (R² 0.91). The time from anchor to release had a mean error of 0.34 seconds, with R² 0.95. A simpler method, which only looks for the moment the string appears in front of the face, was off by about 1.4 seconds on average (R² 0.89): it started the clock during the draw, before the archer had actually reached anchor, and the error grew in archers with a longer draw. A third method, which also uses the arm angle to detect the release, had an error of about 0.512 seconds and R² 0.93, and on the 11 shots where the string could not be picked out it gave a mean error of 0.54. The programme processed more than 30 frames per second on a desktop computer, fast enough to follow the shot live.

The videos were high definition (1080p) at 30 frames per second, filmed from the front with the archer's face and right arm clearly visible; each clip held a single complete shot, from draw to release. The reference is manual timing in Dartfish. The start of the anchor is detected with MediaPipe, Google software that estimates the position of 33 points on the body: from the right shoulder, elbow and wrist the programme calculates the angle of the drawing arm, and treats the anchor as started when the angle is below 25 degrees and has changed on average by less than one degree over the last ten frames.

The release is detected by looking for the string within the box around the face. The image is converted to greyscale and blurred, then the programme searches for edges and straight lines; of the lines found it keeps the one nearest the centre of the face that is tilted between 0 and 30 degrees from vertical, so as not to mistake glasses or the jawline for the string. The release is marked only once the string has been missing for more than ten consecutive frames, so that a momentary loss does not stop the clock. The thresholds (10 for the filter that steadies the points, 25 degrees, 10 frames) were set by the authors through trial and error.

The number of archers is not given: the 58 shots come from archers of various nationalities at a handful of events, and the same archer may appear more than once. The thresholds were hand-picked through trial and error, without saying on which videos, and the programme was not tested on a second set of videos held back for the purpose. Clips in which the arm was heavily obscured or out of frame were removed, and in 11 shots the string could not be picked out: the text does not make clear whether the 0.34-second error is calculated on all 58 or only on the rest.

The method needs footage filmed from the front and a right-handed archer; with a small face in the frame, or lighting that does not separate the string from the background, it gets worse, as the authors themselves say. At 30 frames per second each frame is worth about 0.033 seconds, so a mean error of 0.34 seconds amounts to around ten frames. The study measures nothing about scores: the link between shot time and accuracy comes from other research the authors cite. One of the authors was employed by a company, Techars, as the paper declares.

For anyone who shoots, this is news about tools rather than technique. Timing the stretch from anchor to release currently means going through the video frame by frame; this method does it by itself, from front-on footage, with a mean error of about a third of a second. That is enough to notice that one arrow took two seconds longer than another, not to tell apart differences of a few tenths.

The most interesting finding is why the simple method failed: the string being in front of the face does not mean the archer has anchored: on average the gap was about 1.4 seconds, and it grew in archers with a long draw. Anyone timing their own shots from the moment the string reaches the face is therefore also timing part of the draw. The authors present the programme as an aid to better performance because other studies have linked shorter shot times to higher scores; this study does not measure that.