Scores can be assigned automatically from images of the target face

8 studies · 14 archers measured

Each circle is a study: its position gives the year, its row whether it is for, against or context, its size how many participants it measured. Light circles are studies on other populations or without archers.

In the laboratory, yes; in the field, not yet. Prototypes that read the holes from images work well in clean conditions: the first, with four archers at 10 m indoors, assigns the right score in every sequence tested, and the most accurate, from 2026, locates the holes on a 40 cm face with an error of 1.41 mm and misses the average score per arrow by a median of 1.8 per cent. But the only field test with archers actually shooting, checked against the judges' scores, detects 91 arrows out of 100 and scores 67 of them correctly. The weak points are always the same: arrows that touch or cover another arrow's hole, worn faces, changing light. On torn faces a system trained on smartphone photos drops to around 40 per cent, and even the best one misses about one hole in six, mostly those of tightly grouped arrows, which means the best ends. The two tests with real archers shooting point in opposite directions, and neither was run outdoors at 70 m or over a full round.

Also phrased as: "electronic scoring in archery"; "app that scores from a photo of the target"; "automatic arrow detection"; "camera-based automatic scoring"

In practice

It is taken for granted that apps and cameras on the boss can replace the scorer, and that the only obstacle is price, not reliability.

Evidence

2012
Context General population Published later

10 operators with no shooting experience mark from photographs 72 arrows shot at 70 m by one archer on a 122 cm face: correlation with the measured coordinates of at least 0.977 per axis, agreement between operators ICC of at least 0.969, 4 one-point errors across 720 arrows (8 operators out of 10 made none), group indices within 2 per cent. The largest discrepancies arise where one arrow covers another's hole. It measures human reading of images, not automatic reading.

Inter-Rater Reliability and Criterion Validity of Scatter Diagrams as an Input Method for Marksmanship Analysis: Computerised Notational Analysis for Archery · J. Andrew Callaway, A. Shelley Broomfield
2013
For Archers Published later

4 archers over two days, 6 arrows per trial at 10 m indoors, two cameras at 5 and 10 m from the target, Olympic rules except for the distance: 100 per cent accuracy in every sequence tested, defined as arrows assigned to the right zone out of the total. The overall number of arrows is not reported; no test outdoors or at 70 m.

Image Processing Approach to Automatic Scoring System for Archery Targets · Thi Thi Zin, Ikuo Oka, Takuya Sasayama, Shingo Ata, Hitoshi Watanabe, Hiroshi Sasano
2017
For Model Published later

Camera system close to the target tested on 6 training videos and 14 images: face recognised in 6 videos out of 6, correct score for 29 arrows out of 32 (one lost because it appeared short in the image, one shifted by the paper moving, one lost during a sharp change of light), decimals often underestimated. The test uses the same videos as the calibration. Master's thesis.

Sequential Recognition and Scoring of Archery Shots · Raymond Parag
2021
For Model Published later

Prototype with two cameras on the boss, 16 test arrows: top camera 14 correct (87.5 per cent), side camera 11 (68.75 per cent), combined 14. Two errors attributed to calibration of the side camera; the authors estimate at least 93.75 per cent with correct calibration, an estimate that was not measured. They also report, from a personal communication, that at a US competition judges rule on about 30 arrows out of 18,000. Preprint.

OnTarget: An Electronic Archery Scoring · Andreea Danielescu
2023
Against Model Published later

Deep learning model trained on 480 smartphone photos of 60 cm FITA indoor faces, expanded to 86,400 with rotations and flips: recall and precision of 100 per cent on a test of 5 clean images, around 40 per cent on torn faces. The share of correct scores is not measured, and the authors write that it is unclear whether the system could replace manual counting. No archers.

Automatic Archery Scoring System Using Deep Learning and Image Processing · Haozhe Ma, Michael G. Madden
2023
Context Model Published later

A two-stage system that reads footage of a real shooting session from a tilted camera: on heavily used faces it rebuilds the rings by interpolating their outline better than circle detection with the Hough transform, but the comparison is only pictorial, with no error measure. In the hardest session, with a worn face and overlapping arrows, the scores shown are correct, but an arrow heavily covered by others, in the 9 at about 45 degrees clockwise from east, is placed in the 9 at about 10 degrees clockwise from west. No accuracy percentage is given. Conference proceedings.

Two-Stage Computer Vision Assisted Automatic Archery Scoreboard Scoring System · Jonathan Then Sien Phang, Dar Hung Chiam, King Hann Lim, Basil Andy Lease
2025
Against Archers Published later

10 archers from an Indonesian university group, 10 arrows each, a webcam mounted on the spotting scope 15 m from the target, checked against the judges' scores: 91 arrows detected out of 100, 9 missed, 67 correct scores (a 33 per cent error). The model recognises arrows in 78 per cent of the 850 training and validation images; errors cluster on arrows close to the lines, on overlapping arrows and in poor light. At first use, 90.9 per cent of the 11 respondents needed help. Regional journal.

Real-Time Arrow Detection and Scoring on Archery Targets Using YOLOv8 with Euclidean Distance-Based Zone Estimation · Safri Adam, Novi Aryani Fitri, Sarah Bibi, Muhammad Ridhwan Sufandi
2026
For Model Published later

48 photos of 40 cm indoor faces from a single archer, from 3 to over 100 holes each, 5,084 annotated holes, three-fold cross-validation: F1 0.893 ± 0.011, precision 0.951, recall 0.841, localisation error 1.41 ± 0.06 mm. Error on the average score per image with a median of 1.8 per cent (mean 2.6) and a mean signed error of -0.001 points; group centre within a median of 4.00 mm, against a 20 mm scoring ring. The missed arrows are mostly those within a shaft's width of each other; on heavily worn faces a few phantom holes appear, tears and lifted paper taken for arrow holes. Preprint.

Frozen Vision Transformers for Dense Prediction on Small Datasets: A Case Study in Arrow Localization · Maxwell Shepherd

Status history

2012 Not studied Ten operators with no shooting experience mark by hand, from photographs, 72 arrows shot at 70 m by one archer: the reproduced positions correlate above 0.977 with the measured ones, the operators agree with each other, and across 720 arrows there are four one-point errors, with the largest discrepancies where one arrow covers another's hole. Images are enough to reconstruct the impacts, but a person reads them: nobody has yet tried to have a program assign the score.
2013 Reported A computer vision algorithm recognises the arrows of four archers, shot over two days at 10 m indoors and filmed by two cameras at 5 and 10 m from the target, and assigns the right zone in every sequence tested. It is a single study, at short distance and in controlled conditions, with no test outdoors or at 70 m: the evidence in favour is there and it is limited.
2017 Reported A master's thesis using a camera close to the target recognises the face in six training videos out of six and assigns the right zone to 29 arrows out of 32, often underestimating the decimals. The test runs on the same videos used to calibrate the program, so the figure should be read as an upper bound, and it involves no sample of archers: the status remains reported.
2021 Reported A prototype with two cameras mounted on the boss assigns the right score to 14 arrows out of 16 and shows that position matters most: from above it reads 14 arrows correctly, from the side 11, and combining the two does not help. The errors come from calibration and light, and the benchmark is high, because at a US competition the judges step in on about 30 arrows out of 18,000. A bench test on a handful of arrows, with no archers: the status remains reported.
2023 Reported The first comparison between clean and used faces measures the limit of these systems: a model trained on 480 smartphone photos of 60 cm indoor faces finds every arrow on five clean images, but on torn faces recall and precision fall to around 40 per cent, and the authors write that it is unclear whether it can replace counting by hand. It is a test without archers, so it does not move the status, which stays reported on the strength of the 2013 study. The same year another system, tested on footage of a real session with a worn face, assigns the correct scores in the cases shown but places an arrow covered by others on the opposite side of the 9, without giving an accuracy percentage.
2025 Contested The first field test with archers shooting, checked against the judges' scores, arrives and goes badly: ten Indonesian university archers, ten arrows each, a camera 15 m from the target, 91 arrows detected out of 100 and 67 scores right. The errors cluster on arrows close to the lines, on overlapping arrows and in poor light, and almost every user needed help the first time. With one study on archers in favour and one against, both small, the status becomes contested.
2026 Contested The most accurate validation so far, on 48 photos of 40 cm indoor faces and 5,084 holes, locates the impacts with an error of 1.41 mm and misses the average score per arrow by 1.8 per cent, with the group centre within 4 mm. The error is not symmetrical, though: the system invents few holes (precision 0.951) but misses quite a few (recall 0.841), and the ones it misses are arrows within a shaft's width of each other, exactly where a judge is most needed. These are one archer's targets, photographed outside competition: the status remains contested.

Research lines

102 studiesFrom the total on the scorecard to the arrow's coordinates

Cite

APA

Campagna, M. (2026). Scores can be assigned automatically from images of the target face (Version 3). In Evidence Registry. Retrieved 16 October 2026, from https://matteocampagna.com/en/registry/competition/automatic-scoring-target-images

BibTeX

@misc{campagna2026_punteggioautomaticoimmagini,
  author = {Campagna, Matteo},
  title = {Scores can be assigned automatically from images of the target face},
  year = {2026},
  howpublished = {Evidence Registry, matteocampagna.com},
  url = {https://matteocampagna.com/en/registry/competition/automatic-scoring-target-images},
  note = {Version 3; accessed 2026-10-16}
}

Follow this claim

RSS feed for this claim

RSS is an address you paste into a feed reader: each time this claim changes status, adds a year to its history or is revised, a new item appears there.

Other claims in this domain

See all claims on competition and scoring