Reading the target with a camera

What a target face holds

After an end the score goes on the scorecard, but the position of each arrow is lost as soon as it is pulled from the paper, unless someone photographs it. That is why almost every study of group size, centring and shape goes through an image from which the coordinates of the holes are taken. A person can read it by clicking on the holes, or software can find holes or shafts on its own, reconstruct the scoring rings and assign each arrow to its zone. The networks used to classify and predict are covered on the machine learning page; this page stays with the target.

Photographs read by hand

The study that measures human reading dates from 2012. One archer shot 72 arrows at 70 m on a 122 cm face, each end was photographed and each hole measured to the millimetre; ten operators with no shooting experience then transferred the arrows from the photographs onto a target on a computer. Their coordinates correlated with the measurement at 0.977 or above on each axis, the operators agreed with one another (ICC of at least 0.969), and across 720 readings there were four one-point errors. The largest discrepancies came from an arrow that hid its neighbour's hole in the photograph, a problem that returns in later studies.

The studies of groups at competitions rest on photographs read by hand, or read partly by software and corrected by hand. In Shanghai and Belek in 2013 about 20,000 and 10,000 holes were digitised and the totals checked against the archers' own scores, with overlapping holes and arrows close to the lines assessed by hand. Those photographs produced the 1.15 ratio behind the entry on groups wider than they are tall and the calculation of points lost to an off-centre group.

Away from competition the procedure changes from one study to the next. A Turkish study photographed each face from three angles and measured it in Matlab, and it underpins the entry on arrows landing more often in the upper half; others placed the photographed impacts of 29 archers on a coordinate system, photographed the face from the front and both sides with numbered arrows to recover the order of the holes, scanned faces with the holes marked by adhesive dots, or used an iPhone app or a camera aimed at the target. None of these reports the error of its own reading.

Automatic systems

In the registry the first automatic systems, from 2012, do not use a camera: one puts infrared sensors at the four corners of the target, with correct location in between 60 and 100% of cases according to the author, with no test to show it, and the other works out the impact from the shadow of the shaft on two rows of photodiodes, without reporting any accuracy.

The first camera system to be published came in 2013: four archers at 10 m indoors, the arrow point located where the shaft meets its shadow, and 100% of arrows in the correct zone, out of a number of arrows the paper does not state. A 2017 master's thesis, requested by the Dutch federation, scored 29 arrows out of 32 correctly, but was tested on the same six videos used to tune it. A two-camera prototype, built in 2007 and published in 2021, read 14 arrows out of 16 correctly from above and 11 from the side. A network trained on 480 smartphone photographs found every arrow on five clean faces and about four in ten on torn ones, without measuring the scores, and a system from the same year gives no percentage at all.

The two most recent tests point in different directions. In 2025 a webcam mounted on a spotting scope, 15 m from the target, was checked against the judges on ten arrows from each of ten archers: it detected 91 out of 100 and scored 67 correctly. In 2026 a preprint looked for 5,084 holes in 48 photographs of 40 cm faces: it finds 84.1% of them, 95.1% of its detections match a hole, and the position error is 1.41 mm. The median error on the mean score per photograph is 1.8%, and the group centre falls a median 4.00 mm from the actual one, against a scoring ring 20 mm wide; the photographs, however, come from a single archer. With the laboratory in favour and the only field test against, the entry on scoring from images is "contested", while the one on the errors of camera systems is "reported".

Where the errors come from

The first cause is arrows close together. A hole hidden by another arrow was behind the largest discrepancies even in human reading; in the 2026 preprint the missed holes are mostly those less than one shaft diameter apart, and in the 2023 system an arrow almost hidden by the others stays in the 9 but is placed at a different point on the face. The errors are therefore not random: they cluster in the tightest ends, which are the very groups a study of group size sets out to measure.

The second is the paper: on worn faces, tears and lifted flaps turn into holes that do not exist, and an arrow that moves the paper as it goes in can shift the reading, so much so that in the Dutch thesis a 10 became a 9.85. The third is light, which in 2017 caused one arrow to be discarded and confused the rings in the two-camera prototype. The fourth is perspective: a camera at the side sees the target as an ellipse, which has to be straightened before anything is measured, and in the 2007 prototype that correction is left to future work.

Then there is the line rule. An arrow touching a line takes the higher score, so the thickness of the shaft matters: the 2013 system scores from the point without allowing for it, the Dutch thesis measured from the outer edge of the shaft rather than the inner one, the 2026 preprint subtracts a 2.25 mm radius, and the Shanghai and Belek study, to calculate expected scores, assumes a 5.0 mm shaft, as Archery Australia does. Finally there is tuning: software adjusted on the same images it is tested on tends to perform well on exactly those, which is the case for the Dutch thesis and, in part, for the 2026 preprint, which does not hold back any unseen photographs.

What is missing

No test has been run outdoors at 70 m, over a full round, against the judges' official scores, and the largest of the tests that read arrows one at a time stops at a hundred arrows. The comparison with human error is weak as well: the roughly 30 arrows referred to the judges out of 18,000, reported from a personal communication in the two-camera prototype paper, are the share of doubtful arrows, not a measured error. Where reading the target is missing, studies fall back on something else: the entry stating that two archers with the same score can group very differently rests on an electronic bow aimed at a television screen, so the coordinates come from the simulator rather than from a target face.

What to check

• who reads the image, a person or software, and whether the reading was checked against a measurement on the paper or against the judges;
• where the image was taken from and whether perspective was corrected;
• which point stands for the arrow and how the line rule is applied;
• how overlapping holes, worn faces and arrows off the paper were handled;
• whether the software was tested on images other than those used to tune it, and on how many arrows, at what distance and in what light.

How design and sample size are weighed is explained in How we read a study; what the shape of a group tells an archer is covered in the articles on arrow impact patterns and on why groups open up with distance.

@misc{campagna2026_enresearchmethodsautomatictargetscoring,
  author = {Campagna, Matteo},
  title = {Reading the target with a camera},
  year = {2026},
  howpublished = {Research, matteocampagna.com},
  url = {https://matteocampagna.com/en/research/methods/automatic-target-scoring},
  note = {accessed 2026-10-16}
}