Scoring arrows with a webcam: Adam 2025 study

A group at the Pontianak State Polytechnic, in Indonesia, built a system that finds the arrows sitting in the target and scores them from the image of a webcam mounted behind a spotting scope. A YOLOv8 neural network locates the tip of each arrow; the program then measures the straight-line distance between the centre of the box drawn round the arrow and the centre of the target face, and compares it with the radius of each scoring ring.

In the field test ten archers shot ten arrows each. Out of 100 arrows the system found 91 and gave the correct score to 67; it missed the other 9 altogether. The errors clustered in fading light towards evening, on arrows close to the line between two rings and on arrows that overlapped. In an earlier test on still images, partly taken from the training material itself, the authors report that the model recognised arrows 78% of the time.

On video each frame took 145.05 ms on average, almost all of it (138.11 ms) spent by the network finding the arrows. Archers and judges, given a questionnaire after the test, rated the system as easy to use and said it did not disturb their concentration; 90.9% of them, though, needed help the first time.

The images were collected with the support of the polytechnic's archery club. A first attempt with the sensor 3 metres from the target risked having it hit by an arrow, and a smartphone produced files that were too large, so in the end the webcam was fixed to a spotting scope on a tripod, whose magnification also gave a sharper picture. To the photos they took the authors added images from Kaggle, for a total of 850, split 80 to 20% between training and validation, each with a single hand-drawn label, "arrow", around the tip. Training ran for 100 epochs on Google Colab, with a few random alterations of the images (blur, median filter, greyscale, contrast) each applied with a probability of 1%.

The screen shows guide circles worked out from the sizes of the World Archery target faces of 122, 80, 60 and 40 cm, which have to be lined up with the real target by adjusting the camera's distance and zoom; in the test the camera stood 15 metres from the target. The automatic score was compared arrow by arrow with the one recorded by the judges. The paper does not give the shooting distance, the type of bow or the size of the target face used in the test.

The main result rests on 100 arrows from ten archers at a single site, and the paper's figures do not always agree with one another. The method section describes a 70 to 30 split of the data, the results section an 80 to 20 split; the images collected are 200 in one place and 650 in another. The 33% is called a detection error, yet only 9 arrows were missed and the other 24 were found but scored differently from the judges. The evaluation formulas also refer to shooting actions detected or missed, as if they came from another project, and it is not clear which data the 78% was calculated on.

The score depends on how well the guide circles are aligned: the authors themselves write that small movements of the scope were enough to cause significant scoring errors, and the system has no automatic calibration. An arrow on the line, the case that matters most in competition, is exactly the one the system handles worst, and there is no measurement of the position error in millimetres. The questionnaire had few respondents (the percentages suggest 11) and says nothing about accuracy. In the comparison the authors themselves report, several other systems, not all designed for archery, exceed 90%.

For anyone shooting today the study changes nothing: a system that scores two arrows out of three correctly cannot replace someone scoring by hand, in competition or in training. It is more useful for seeing where the difficulty of automatic scoring lies: finding the arrow in the target is the easier part, while placing it in the correct ring when it touches the line, in fading light or next to another arrow, is the hard part.

For anyone thinking of recording their own arrows to study their groups, the paper shows that a webcam behind a spotting scope is enough to see the arrows from 15 metres and that the position of each impact can be extracted automatically. It does not show that this position is precise enough to measure an archer's spread, because the authors only check whether the ring matches the one recorded by the judges.