An overly short grace treats brief occlusion as deliberate disengagement

What it is

Tracking drops for a couple of hundred milliseconds—the other hand blocks the view, a collar enters the frame, a headset slips—and the user’s intent never left. If grace is shorter than those drops, the system writes the occlusion as deliberate disengagement, clears engagement, and interrupts the in-flight action. The user must perform the enter action again to continue. The felt experience is “I was still holding; it left on its own.”

Why it happens

The duration distribution of short occlusions is physical: a torso walking past is often 300–800 ms, self-occlusion (one hand paging in front of the other) can be tens of milliseconds, looking aside at a person is about half a second. Grace set as “exit when a frame has no hand” treats the sensor’s momentary failure as a user decision. Once engagement is cleared, the next appearance of the hand is a new arrival: focus is gone, the selection is gone, the enter animation replays. What breaks is session continuity, not the recognizer’s class label. Continuous tasks (scrubbing a timeline, holding a pinch to set volume) are especially sensitive, because the body is being used as a key held down.

Studying it

Insert occlusions of known duration during engagement, covering self-occlusion, clothing, and passers-by. Measure whether the action is marked disengaged, whether the user must repeat the enter action, and time from occlusion end to being able to work again. Have participants mark “I thought I was still operating.” Sweep grace from 0 to 1 s and find the inflection where false exits drop. Separately record whether the user still held the operating pose during the occlusion—still posed, yet exited, is the signature of too-short grace.

Where it stops holding

On a tap-like exhibit where people exit often anyway, the cost of a false exit is one extra enter, often acceptable. Flows that control a moving object or are expensive to restart (alignment, measurement, recording) cannot accept it. Infrared or stereo drop more fragmented frames in the dark, so the same grace that suffices in a bright lab looks too short under gallery lighting. Headset hand tracking and desktop cameras have different loss statistics; they cannot share one millisecond number.

Applying it

• Set the lower bound from measured occlusion-duration percentiles in the venue, so common self-occlusion falls inside the window.
• When a false exit happens, do not only return to an idle desktop: if an in-flight operation can be inferred, prompt “continue what you were doing?” after recovery, rather than pretending it never started.
• Accept with two scripts—“one hand pages while the other is blocked” and “someone walks in front of the lens”—and count forced repeats of the enter action.

Related

• Same group: C4.19.1 A re-entry grace period can keep engagement after brief tracking loss · C4.19.2 An overly long grace keeps the system live after the user has abandoned the action · C4.19.4 Whether re-entry requires repeating the enter action depends on whether the grace has expired
• Adjacent: C4.23 Handling tracking loss · C4.31 Lighting, occlusion, and environmental conditions
• Search: premature disengagement · occlusion · self-occlusion