A camera-based scoring system misreads some arrows: the ones barely sticking out and the ones that shift the target face

The summary today

Every prototype tested makes errors, and their causes are described well in only one paper. A 2017 thesis, written at the request of the Dutch federation, tested a computer vision system on six training videos: 29 arrows out of 32 were given the correct zone, and the three misses had three different causes. One arrow shifted the target face as it went in and the program read a ten as 9.85; another stuck out too little in the image to be recognised as a shaft, and lowering the threshold makes arrows appear that are not there; a third, arriving during a sharp change in light, was discarded. The program had, however, been tuned on those same six videos. Later tests confirm that errors happen, with other causes: a two-camera prototype read 14 arrows out of 16 correctly from above and 11 from the side, and the author blames the side camera's calibration; in 2025 a system tested in the field on 100 shots by 10 archers failed to see 9 arrows and gave the correct score to 67, and the authors point to light, arrows close to the line and overlapping arrows; and systems that count holes from photographs mostly miss the overlapping holes of the tightest ends and take tears in worn faces for holes. The tests on arrows shot and read one by one go no further than a hundred arrows, the hole counting from photographs comes from a single archer, and no test compares the system with the judges during a competition.

Also phrased as: "electronic scoring"; "app that scores from a photo"; "computer vision archery"; "automatic target reading"; "camera on the target"

Evidence

• 2012
  Context Equipment tests Published later

  Infrared sensor system mounted on the target, not camera-based: location accuracy between 60 and 100% on the 8 × 8 grid, three arrows read in about 3 s, 80 cm face at 30 m.

  Approach to Improve Archery Competence through Creative Engineering Practice · Hajime Miyauchi

• 2012
  Context Equipment tests Published later

  A system for quality control of arrows in the factory, not camera-based: two line lasers and two rows of photodiodes in a 65 × 65 cm frame about a metre in front of the target work out the point of impact from the shadow of the shaft, using fuzzy logic that avoids the image correction cameras need. Arrows fired from a machine at about 275 km/h. It reports neither an accuracy figure nor any misread arrows. Regional journal.

  Measuring System for Impact Point of Arrow using Mamdani Fuzzy Inference System · Jung-Won Yu, Han-Soo Lee, Yeong-Sang Jeong, Sung-Shin Kim

• 2017
  For Equipment tests Published later

  Computer vision system with a camera set at an angle near the target, tested on 6 training videos from the Dutch federation: 29 arrows out of 32 in the correct zone (7/7, 6/6, 1/2, 5/5, 6/6, 4/6). One arrow shifted the face as it went in and was read as 9.85 instead of ten; one stuck out too little in the image to be seen, and lowering the threshold makes non-existent arrows appear; one, arriving during a sharp change in light added to the video, was discarded because the system detected more than 20 lines. The program was tested on the same videos used to tune it; master's thesis.

  Sequential Recognition and Scoring of Archery Shots · Raymond Parag

• 2021
  Context Equipment tests Published later

  Two-camera prototype on a frame around the boss, 16 test arrows: top camera 14 correct (87.5%), side camera 11 (68.75%), combined 14. The author puts the errors down to the side camera's calibration, as well as focus and changing light, and estimates at least 93.75% once it is corrected. Preprint.

  OnTarget: An Electronic Archery Scoring · Andreea Danielescu

• 2023
  Context Equipment tests Published later

  Score reading from smartphone photos of 60 cm indoor faces: recall and precision around 40% on randomly torn faces, 100% on a separate test of 5 photos of clean faces; dataset of 480 original photos. No figure for how many arrows receive the correct score.

  Automatic Archery Scoring System Using Deep Learning and Image Processing · Haozhe Ma, Michael G. Madden

• 2025
  Context Equipment tests Published later

  YOLOv8 system with a calibrated camera 15 m from the target, tested in the field against the judges' scores: 10 archers from an Indonesian university group, 10 arrows each; 9 arrows out of 100 not detected, 91 detected, 67 correct scores (67% accuracy). The authors put the errors down, without counting them arrow by arrow, to poor light towards dusk, arrows close to the line between two zones, overlapping arrows or positions missing from the training images, and calibration: neither of the two causes named in the claim. 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
  Context Equipment tests Published later

  Hole counting on 48 photos of 40 cm indoor faces (5084 holes marked by hand), three-fold cross-validation: precision 0.951 ± 0.009, recall 0.841 ± 0.018, mean location error 1.41 mm. The missed holes are mostly the overlapping ones of tightly grouped arrows; holes reported where there are none come from tears and lifted paper on worn faces. Median error on the mean score per photo 1.8%. A single archer and a single range; preprint.

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

Status history

1. 2012 Not studied The first automatic systems in the bibliography do not use a camera. One puts infrared sensors on the target and estimates the position and score of the arrows, with a location accuracy between 60 and 100% on an eight by eight grid, reading three arrows in about three seconds; another, designed to check arrows in the factory, works out the point of impact from the shadow of the shaft on two rows of photodiodes lit by lasers, and rules out cameras precisely to avoid having to correct the images. The first shows that automatic reading still makes mistakes, but nobody has tested a camera system or described which arrows slip through.
2. 2017 Reported A computer vision system tested on six training videos from the Dutch federation assigns the correct zone to 29 arrows out of 32. The errors have specific causes: the target face moving around the hole, so that a ten was read as 9.85; the barely protruding arrow that goes unseen, with the opposite risk of making non-existent arrows appear if the threshold is lowered; and a jump in brightness that caused one arrow to be discarded. It is a single test, on a few videos that were also used to tune the program: the supporting evidence is there and it is limited.
3. 2021 Reported A prototype with two cameras around the boss reads 14 arrows out of 16 correctly with the top camera and 11 with the side one, and the author puts the errors down to the calibration of the second, as well as focus and changing light. It confirms that the system misreads some arrows, but for a different cause from those of 2017 and on 16 arrows: the picture stays as it was.
4. 2023 Reported A system that reads the score from a smartphone photo finds every arrow on five photos of clean faces, but on torn faces recall and precision fall to around 40%. It adds worn paper to the causes of error, without measuring how many arrows receive the correct score: the status does not change.
5. 2025 Reported The first field test against the judges' scores arrives: ten Indonesian university archers shoot ten arrows each with a camera 15 m from the target, and the system fails to see 9 arrows out of 100 and gives the correct score to 67. It confirms that automatic reading loses arrows, and more often than in 2017, but the errors are put down in words, without counting them, to poor light, arrows close to the line, overlapping arrows and calibration, not to barely protruding arrows or a moving face, and the journal is regional: on the causes named in the claim there is still a single study, and the status does not change.
6. 2026 Reported A system that counts holes from photographs of 40 cm indoor faces finds 84% of them and invents few, with 95% precision: it mostly misses overlapping holes, in other words the tightest ends, and on badly worn faces it takes tears and flaps of paper for holes. On the mean score of each photo the median error is 1.8%. It is a third family of errors, from photos of a single archer at a single range: the picture stays as it was.

Related

Related articles

• The future of archery: technology, research, evolutionWhat has already been measured about technology in archery: virtual reality, sensors, video, the quiet eye, camera scoring and shooting machines.

Research lines

• 102 studiesFrom the total on the scorecard to the arrow's coordinates
• 35 studiesTeaching and learning archery: sixty years of instructional trials

Other claims in this domain

• CorroboratedAlong its flight the arrow loses an appreciable part of its launch speed12 studiesStatus since 2011
• CorroboratedAs it leaves the bow, the arrow flexes and snakes around the riser instead of leaving straight19 studies · 47 archers measuredStatus since 2012
• ReportedA chest band with fibre optic sensors measures breathing and mean heart rate with small errors, and an archer can wear it while shooting2 studiesStatus since 2019
• ReportedA compound bow stores between 70 and 100 joules and sheds about two thirds of the force at full draw4 studiesStatus since 2015
• ReportedA damper tuned to the tremor frequency reduces the bow's vibration while aiming1 studyStatus since 2017

See all claims on equipment and tuning