Quiet eye in archers via neural networks: Söğüt 2025

Quiet eye is the last fixation of the gaze before a movement begins, and it is usually measured by having an expert annotate, frame by frame or stretch by stretch, where it starts and where it ends. Five researchers at Mersin University tried handing that work to a neural network, starting not from an eye tracker but from electrooculography, the electrical potential measured around the eye.

They recorded ten licensed archers from a club in Mersin, six arrows each, for a total of sixty stretches of signal labelled by hand by an expert. On those sixty stretches they trained and compared five architectures: a one-dimensional convolutional network, a Transformer, a UNet, a convolutional network coupled with a GRU and a convolutional network coupled with an LSTM.

In the results table the convolutional network with LSTM comes out best, with an accuracy of about 93%, a positive predictive value of 0.948, a sensitivity of 0.962 and an F1 score of 0.954. Next comes the convolutional network with GRU at about 92% (F1 0.928), then the UNet at 91% (0.875), the plain convolutional network at 90% (0.852) and the Transformer at 89% (0.847), the last of these with the widest spread on positive predictive value. The abstract of the same paper instead reports 95% and 93% for the first two: those are the numbers in circulation, but they are not the ones in the table, and this record follows the table.

The study measures the ability to recognise quiet eye, not its relationship with shooting. There are no scores, no arrows counted, no comparison between quiet eye duration and outcome: the yardstick is the expert's annotation, and the result is how well a network reproduces it.

Eye movements were recorded with silver and silver chloride disc electrodes around the eye, on a BIOPAC MP100 system, sampling at 200 hertz. Since the amplifier was single-channel, the horizontal and vertical recordings were made separately. Quiet eye duration is defined as the time from the start of fixation to the release of the arrow, and its end point was annotated by hand by an expert on the basis of the sound of the arrow leaving the bow.

The signal was cleaned in two steps. First a ten-level wavelet decomposition, since the tenth approximation coefficient corresponds to the drift of the baseline and subtracting it gives a centred signal; then a Butterworth band pass filter over the electrooculography range, between 0.1 and 20 hertz. For each arrow the usable stretch was trimmed from 2.5 seconds before to 6 seconds after the quiet eye duration, giving sequences of 1,700 samples, and labelled in binary, with 1 for quiet eye stretches and 0 elsewhere. To make training more robust the stretches were multiplied by time shifting and adding Gaussian noise.

The five networks were trained for fifty epochs with the Adam optimiser and a learning rate of 0.001, and assessed on accuracy, positive predictive value, sensitivity and F1 score. The model the authors propose has two convolutional layers of 64 and 128 filters, three bidirectional LSTM layers of 200, 100 and 50 units with L2 regularisation, upsampling layers to restore the sequence to its original length, and a final layer with a sigmoid function producing the classification. Preprocessing is done in MATLAB, the networks in Python with TensorFlow and Keras.

What is known about the ten archers is that they are licensed, that they have been training for at least two years, four days a week for two hours, and that they average 160.9 centimetres and 63.7 kilos. Ethical approval comes from Osmaniye Korkut Ata University, and parental consent was collected for the participants under eighteen.

The most important thing to know is that this record is written on the preprint, deposited on 9 June 2025, while the registry entry points to the version published in a journal: the two texts may differ, and the numbers given here are the preprint's.

The sample is ten archers and sixty arrows in all, and the authors themselves write that ten athletes are enough for a preliminary investigation but do not cover the range of technical levels or signal shapes. Almost nothing is known about those ten: not their age (only that some were under eighteen, from the parental consent), not their sex, not their discipline, not their level, and not the distance they shot. Six arrows each is little even for describing one person's quiet eye.

On the training side the decisive point is missing. The paper does not say how the material was split between training and validation, and in particular does not say whether data from the same archer, or copies produced by shifting and adding noise to the same stretch, ended up on both sides. With sixty stretches enlarged by augmentation, an accuracy of 93% on a validation set not separated by participant says nothing about how the method would do on a new archer.

Then there is the question of what the reference is. The networks are trained to reproduce one expert's annotation, not an independent measure of quiet eye: if the annotation is off, the network learns to be off in the same way, and there are no two annotators to compare agreement between. The end of quiet eye is taken from the sound of the arrow, which is a workable criterion but a coarse one against the milliseconds under discussion. The figure numbering is confused: number 10 appears twice and the captions do not match what the text says. The preprint is not peer reviewed, and the study had no funding.

For someone who shoots, nothing here is guidance about shooting. The study does not say how long quiet eye lasts in archers, does not say whether a longer quiet eye goes with better scores, and does not compare different levels: it does one thing only, which is to show that a neural network can find, inside an electrooculographic signal, the stretches an expert had marked as quiet eye.

Its value is as a tool. Measuring quiet eye by hand is slow and depends on who is looking, and that is the bottleneck keeping studies on the subject small: if the marking can be automated, it becomes possible to analyse hundreds of arrows instead of dozens. The authors say so plainly, and talk about a future real-time feedback system, which does not exist here: the method works on recordings already made.

The choice of starting instrument is interesting. The authors use electrooculography rather than a portable eye tracker because, they write, the latter often samples at 30 hertz, while rapid eye movements need at least 200 hertz to avoid distortion. That is an observation about instrumentation worth noting for anyone thinking of measuring their own gaze with a commercial system: temporal resolution matters, and the durations discussed under quiet eye run from a few tenths of a second to a few seconds.