Train/test split should be done manually, with no shuffling, and break up dataset on the basis of 1) holding out 1-2 users, 2) holding out half of the gestures for 1-2 (other) users, and 3) holding out half of each gesture (or rather half of the trials for a given gesture) for 1-2 users.
Which users are chosen for the test split probably will greatly influence the results......
Train/test split should be done manually, with no shuffling, and break up dataset on the basis of 1) holding out 1-2 users, 2) holding out half of the gestures for 1-2 (other) users, and 3) holding out half of each gesture (or rather half of the trials for a given gesture) for 1-2 users.
Which users are chosen for the test split probably will greatly influence the results......
Also include a val split? ...