Gestures can be continuous (for example, running, swimming) and non-continuous, which have a start and finish (for example, swipe left, swipe right, jump). Non-continuous gestures require additional data preparation step: segmentation of raw data into perfect samples with gesture signal peak being in the middle of the feature extraction window. You must (subjectively) decide what is the timeframe of the longest gesture (for example, 1 second), then given the sampling rate (for example, 100Hz). This will be the window for all gestures segmentation in the raw data in order to create perfect training samples.
To simplify this process, you can use a signal centering script that automates the segmentation step.
Note
The provided script performs this data-preparation step automatically when you supply the timeframe value representing the duration of the longest gesture.