Recognising shot phases from video: Muslim 2025 study

The authors, computer scientists at Telkom University in Indonesia, tried to get a computer to recognise which phase of the shot an archer on video is in. There were three phases: set-up, draw and aim, and expansion and release. They used videos of 20 recurve and barebow archers, some taken from Instagram and YouTube and some filmed in the field, extracted joint positions with YOLOv8, a pose estimation programme, and trained two classifiers, a support vector machine and a random forest.

On the test data the support vector machine correctly classified 90% of frames for recurve and 76% for barebow, and the random forest 87% and 75%. The random forest reached 100% on the training data, a sign that it had fitted those data too closely. Errors were concentrated between aiming and expansion: for barebow the support vector machine recognised only 39% of draw-and-aim frames, and the random forest 58% of expansion-and-release frames. The authors put this down to the similarity between frames from neighbouring phases and, in barebow, to differing release styles, with the hand staying near the face or ending up behind the head.

Each video was cut into clips of about two seconds per phase, brought to a common frame rate and broken down into at most 50 frames. On each frame YOLOv8 locates 17 body points (nose, eyes, ears, shoulders, elbows, wrists, hips, knees and ankles); some were discarded because they do not move during the shot or are not visible. The point coordinates, labelled by phase, were split 75% for training the classifiers and 25% for testing them, separately for recurve and barebow. The support vector machine's parameters were chosen by grid search, while the random forest used 100 trees. Performance is reported as accuracy, precision, recall and F1, with confusion matrices and learning curves.

The weakest point is how the data were split: the paper does not say that archers were kept apart between training and testing, so near-identical frames from the same shot probably ended up in both sets, which inflates accuracy. Nor does it say how many of the 20 archers shot recurve and how many barebow, what level they were, how many shots each contributed, or whether the people in the videos taken from social media gave consent.

The phases themselves are loosely defined, and cutting the two-second clips by hand decides where aiming ends and expansion begins: part of the confusion between the two starts there. The study classifies frames rather than assessing technique, and measures nothing to do with shot quality or score. Finally, the text sometimes cites the figures out of step with their content (the learning curves and feature importance are credited to figures other than the ones that show them).

For archers, this work changes nothing for now: it is a first step towards tools that recognise the phases of the shot in a video on their own, and it is mainly of interest to people developing analysis software. Recognising the phase is not the same as assessing the movement: a system like this can tell whether the archer is aiming or has already released, not whether the shot was a good one.

The most useful point for coaches is indirect. The computer struggles most to tell aiming from expansion, because from the outside the difference in position is tiny, and with barebow, where release styles vary a lot from one archer to another. For tools of this kind, the move from aiming into expansion remains the hardest point to read in ordinary video.