Deep learning on archery shot patterns: Park 2025 study

Four Korean national-level archers, all men, were filmed in training between February and May 2024. From each shot, from the set-up position to release, pose estimation software extracted joint coordinates frame by frame; of the 6,231 shots recorded, 5,984 remained, those under 900 frames, that is under 15 seconds. Several sequence neural networks were trained to recognise which of the four archers was shooting, using only the trajectories of the shoulders, elbows and wrists.

The best, a bidirectional GRU, identified the archer in 98.7% of cases on the test data; the plain GRU reached 98.5%, the bidirectional LSTM 96.6%, the simple recurrent network 89.7% and the plain LSTM only 55.9%, a result the authors put down to overfitting or insufficient tuning. The system then shows, on graphs and overlaid on the video, which joints in a shot depart most from that archer's usual movement, highlighting those beyond an error threshold of 0.4.

The videos were Full HD at 60 frames per second, with the whole archer in frame. Pose was estimated with RTMO, a system that locates 17 body points following the COCO standard; the authors dropped the 11 points for the head, hips, knees and ankles, regarded as stationary during the shot, and kept the 6 for the shoulders, elbows and wrists, with their x and y coordinates. Coordinates were normalised to compensate for differences in camera angle and distance from one day to the next, and points estimated with low confidence were reconstructed from the five frames before and after.

The networks had two recurrent layers (128 and 32 units), a layer averaging over the whole sequence and two final layers; hyperparameters were searched with Optuna, and training stopped when the validation loss had not improved for 10 epochs. The data are published in an online repository. The paper does not say how the shots were split between training and testing.

The task the model solves, recognising who is shooting among four people, is much easier than what the system promises, which is to say whether a shot is good or consistent. Four archers can differ in build, position relative to the camera and filming days, and a network can tell them apart partly on that basis. Without knowing how the data were split, shots from the same session, almost identical, may have ended up in both the training and the test data, inflating the accuracy. The paper also contains a technical contradiction: it gives both RMSE and cross-entropy as the loss function.

Above all, no measure is linked to score or to where the arrows landed: that the joints furthest from the usual movement are the ones to correct is an assumption. The authors write that coaches were able to give more precise feedback and that the athletes improved technically as a result, but the study does not measure this. The analysis is two-dimensional, the bow type is not stated and there are only four archers.

The study shows something that coaches at high level already sense: even reduced to the movement of shoulders, elbows and wrists, each archer's shot appears recognisable, so much so that a network identifies it almost every time, although it may partly be relying on build and camera framing. This supports comparing each shot with the same archer's usual movement rather than with an ideal model shared by everyone.

The rest is a promise, not a finding. The system flags how far a shot departs from that archer's average, but it cannot tell whether the departure makes the shot worse, and it does not distinguish a harmful variation from a harmless one. For an archer filming with a phone, the useful part is the principle rather than the tool: shots filmed from the same spot every time make the differences between shots visible, but what they mean still has to be linked to where the arrows land, which this study does not do.