The capture indicator should be independent of recognition results and remain on even when no gesture is detected

What it is

The indicator follows “capture is on,” not “this frame had a gesture.” With nobody waving and the classifier emitting “none,” the lamp should still mean capturing. Wire the indicator to recognition results and it goes dark when idle and lights when busy; users learn “dark = not looking,” which is the opposite of the truth. Independence is about the signal source: power or the data stream, not the classifier’s argmax.

Why it happens

Recognition results are sparse events; capture is a dense process. Driving a lamp that should represent a dense process from sparse events makes duty cycle follow user activity, not privacy state. A common implementation error is reusing “gesture cursor visibility” as the capture indicator: the cursor hides when there is no hand, and the indicator vanishes with it. Another is driving lamp brightness from confidence, so a low-confidence stare looks like not capturing. Once independent, recognition success uses a second cue (sound, animation) and the capture lamp stays steady. Then “it is still looking” in the idle and “that one counted” in the busy do not fight over one pixel.

Studying it

Record a long idle with a few gestures in it, and plot indicator brightness against recognition events. If the two curves coincide, the indicator is not independent. Have participants report whether capture is on in idle stretches and in gesture stretches separately. Compare wiring to power, to “a hand is present,” and to “a valid gesture.” Collect privacy ratings in the idle on its own; that is the interval of “not recognized, still seen.”

Where it stops holding

A power-saving policy that kills the sensor after confirming nobody is there should have the indicator follow power: dark means truly not capturing, which still counts as independent—the cause of dark is power, not “nothing was recognized.” A sensor still warming up has no frames yet and should show “starting,” not pretend it is already capturing. With several sensors, the indicator must say which path is capturing, or radar still on with RGB off will be misread. A debug overlay is not an end-user independent indicator.

Applying it

• The indicator’s input is a liveness signal from the capture pipeline. In code review, forbid wiring it to the gesture callback.
• Idle acceptance with no gesture: the lamp or icon must still look “on”; gesture success gets a separate brief cue.
• Make a wiring table with columns power / hand present / gesture present, and tick only power. A wrong tick means change the implementation, not explain in copy why the lamp goes out.

Related

• Same group: C4.29.1 Whether the sensor is currently capturing should have a persistent visible indicator, not only a cue at recognition time · C4.29.3 Without capture-state visibility, users cannot tell whether silence means not capturing or not recognizing · C4.29.4 Turning capture off should be as reachable as turning it on
• Adjacent: C4.24 Recognition confidence and bias direction · C4.26 Distinguishing three-state feedback
• Search: independent indicator · idle capture · not recognition-gated