Figure 12.

Confusion matrices of a shallow DeepConvLSTM applied to the Hang-Time dataset. The model was trained with a 1-layered recurrent part with 1024 hidden units, a sliding window of 1 s with $50\%$ overlap, and a fixed random seed. The left confusion matrix (a) is obtained from averaging the per-subject LOSO results using the drill and warm-up data as input data. The right confusion matrix (b) is obtained from training on the drill and warm-up data and validating on the game data. One can see an increase in overall confusion when applying the architecture to in-game data, i.e., data recorded in an uncontrolled environment.