Scoring arrows from a photo: Ma 2023 study

Two computer scientists at the University of Galway propose a system that reads the score from a smartphone photo of the target. The work has two parts. A neural network (a modified U-Net, borrowed from a method designed for counting people in crowds) looks for the point where each arrow enters the paper; an image-processing stage recognises the target face from the colours of its rings, corrects the perspective, finds the centre and gives each arrow the value of the ring it lands in.

The results are few and the authors present them cautiously. On a validation set made of randomly torn target faces, recall and precision were around 40%: the network found roughly four arrows in ten, and roughly four of its detections in ten were arrows. On five photos of clean faces it found every arrow, with no extra detections. The authors themselves say they cannot tell whether this architecture will ever be accurate enough to replace reading and counting by hand. No figure is given for the final score, meaning how many arrows end up with the correct value.

To train the network the authors photographed 480 targets during practice, with a range of scores, arrow colours, angles, backgrounds and degrees of paper wear. Each arrow was marked by hand with a point and its value, from 1 to 10, plus a class for misses; rotating and flipping the images produced another 86,400. These are FITA indoor faces, and the name of the dataset published on Kaggle points to the 60 cm face.

The network was trained for about 1,000 epochs, until the loss fell below 10, and then tested on the two sets of images described above. The scoring stage isolates the face with a colour mask (on the blue, for instance), reduces noise, corrects the skew of the rings, works out their dimensions and matches the network's coordinates against them. The text recalls that an arrow touching a line takes the higher score, but does not explain how the system applies that rule.

This is a short contribution, four pages presented at a conference, describing a prototype partway through development. The testing is minimal: the size of the validation set is not given, and the successful test covers five photos. We do not know how many arrows they contained, at what distance or with which phone they were taken, or whether they came from the same sessions used for training. The extra 86,400 images are rotated and flipped versions of the same 480, not new photos.

Above all, the measure that matters to an archer is missing: how many arrows receive the correct score, and how far the estimated point lies from the real one. Turning this into an app that works on a phone is, in the authors' words, future work.

For anyone shooting today nothing changes: the system is not ready, and its authors say so. The work does help to show where the difficulty lies in counting points from a photo. On fresh paper the holes are easy to find; on a worn face, with tears and old holes, the network got confused and recognised fewer than half of them. An app that reads scores through the camera may run into a similar limitation at the end of a session, when the face is full of holes, but the study tests no commercial app. In the registry it counts as the description of a tool, not as evidence that a tool works.