Randomly sample some number of template gestures (training data) and then leave the rest out as the testing data.
$B was a little different since AFAIK it didnt do a train/test split (did some kind of crossval / hold one user/template out) and was just matching incoming data to data already in the training set (doing pairwise comparison for every gesture and only works for gestures for which a template exists in the training dataset)
Randomly sample some number of template gestures (training data) and then leave the rest out as the testing data.
$B was a little different since AFAIK it didnt do a train/test split (did some kind of crossval / hold one user/template out) and was just matching incoming data to data already in the training set (doing pairwise comparison for every gesture and only works for gestures for which a template exists in the training dataset)