McNeill's taxonomy classifies gestures by what they do rather than what they look like, which is why it survives translation to image coding: an iconic gesture depicts content, a deictic points, a beat marks rhythm, and an adaptor manages the speaker's own state rather than communicating at all.
Intended for detection studies. Draw a box around the hand or hands performing the gesture — not the whole person — and code its function. Two hands doing different things get two boxes.
The hard limit of frame coding. Gesture is a movement, and a single frame shows a posture. An iconic gesture caught at rest looks like nothing; a beat is invisible without its stroke. Code from frames only what the frame really shows, and prefer a video timeline when the distinction between beat and iconic matters. This is the scale where the honest answer is most often "Partial".
Adaptors are the clinically interesting code. Self-touch — face touching, neck rubbing, hand wringing — is associated with discomfort and cognitive load, and it is the one category that is not addressed to the listener. Coding it separately keeps it out of the communicative-gesture count.
Severity is clarity. If the hand is blurred mid-stroke or leaving frame, that is Partial or Uncertain no matter how confident your reading of the context is.