Arrow holes located by computer vision: Shepherd 2026

The paper presents a system that, from a photo of a 40 cm indoor face, finds every arrow hole, measures its position relative to the centre and calculates the score under the line-cutting rule. The author, from Johns Hopkins University, trained it on 48 photos with 5,084 holes marked by hand and evaluated it with three-fold cross-validation, each time testing the model on photos it had not learned from. On average the system found 84.1% of the holes (recall 0.841), 95.1% of its detections were holes (precision 0.951), giving an F1 of 0.893 and a mean position error of 1.41 mm.

For the mean score per photo, the median gap from the score worked out from the hand-marked holes is 1.8% (mean 2.6%), with no tendency to over- or underestimate. The centre of the group lands a median of 4.00 mm from the real one (mean 4.90 mm), less than a quarter of the width of a ring, which on the 40 cm face is 20 mm. The holes it misses are mostly overlapping ones, where two or more arrows go in less than a shaft diameter apart; detections with no hole behind them come from tears and flaps of paper on heavily worn faces.

The 48 photos were taken during practice, from different angles, with anything from 3 to more than a hundred holes. A first stage of classical computer vision straightens the image: it recognises from the colours the boundaries between yellow and red, red and blue, blue and black, and black and white (at 40, 80, 120 and 160 mm from the centre), fits ellipses to them and warps the photo until the rings become concentric circles, at a scale of about 0.195 mm per pixel.

This image is then processed by a large vision network already trained on general images (DINOv3), which is kept frozen; the author adds small trainable layers, 3.8 million parameters out of 308 million, which produce a map of hole centres. To make up for the small number of photos, the images were varied with rotations, flips, blur, noise and changes in lighting. A detected hole counts as correct if it lies within 15 pixels, or 11.7 mm, of the hand-marked one. For scoring, the distance from the centre is reduced by the shaft radius (2.25 mm, for 4.5 mm shafts) so as to apply the line-cutting rule.

The author lists the main limitations himself: the 48 photos come from a single archer at a single venue, and the system was tested only on the 40 cm face. There is no set of photos held back and never seen during model selection: the settings were chosen and then evaluated on the same three folds, which may make the results look slightly better than they would on new photos.

The 11.7 mm criterion for counting a hole as found is generous compared with a 20 mm ring, although the mean error on the holes found is still 1.41 mm. One detail does not add up: the detection threshold chosen is 0.5, while the search described explored values from 0.01 to 0.4. The comparison with an earlier system for outdoor targets (1.77 mm error) is not direct, as the author admits, because the faces, images and criteria differ. It is a preprint on arXiv, not yet peer reviewed.

The study is not about shooting but about a tool that could be useful to archers: getting from a photo of the face not just the total but the position of every arrow, and so the grouping and the average offset of the group from the centre. On photos from a single archer the system recovers the mean score to within a few per cent and the group centre to within a few millimetres, enough to see whether the group consistently sits to one side. Its weak point is precisely the face of someone who groups well: when arrows pile up in the same spot and the holes overlap, the system counts fewer than there are. How it behaves on other faces, in other light and with other phones, has yet to be tested.