The problem

A cue that is correct in content and late in delivery has failed.

In small theatre, the person who knows when a cue belongs and the person who can actually fire it are almost never the same. An assistant director watches from the wings with nothing to press. A volunteer runs the lighting desk off a cue sheet. So every cue has to travel between them: seen, signalled, understood, found on the desk, pressed. Each handover costs time, and the lights land late, night after night.

The idea

  1. Perceive the moment
  2. Signal the operator
  3. Operator understands
  4. Find the control
  5. Execute
Manual cue delivery: five handoffsGesture trigger: three removed
Handing the trigger to the person who already holds the knowledge removes three of the five stages.

The five cues

The prototype reads five cues from your bare hands — no markers, no gloves, nothing to calibrate. The set is deliberately tiny, and picked so the gestures are hard to mistake for each other: some held still, some moving, some one hand, some two.

Hover a cue and the lights do it to him.

CueGestureKind
BlackoutDouble clapDynamic · two hands
General coverBoth palms open, raisedStatic · two hands
Centre specialIndex finger pointing, heldStatic · one hand
Master dimFlat palm height sets the levelContinuous · one hand
Colour stateLateral swipeDynamic · one hand

What we actually measure

Papers on gesture recognition routinely report accuracy in the nineties. Those numbers come from short clips that were cut to contain a gesture in the first place, so they cannot measure false firing at all. On a stage the two errors are not equal: a missed cue just fails to help, but a false one blacks out the stage in front of an audience.

So the headline number is false triggers per hour, alongside cue delivery latency against a manual baseline. Cues are grouped by timing contract: timing-critical cues commit as soon as they are recognised, while selective ones wait behind an explicit confirmation. Three detectors run on identical data through an identical streaming loop: naive rules, rules with structural fixes, and a small causal temporal network. Evaluation is split by participant, not by clip.

There are no published results yet. I will put them here once the testing is done.

Reading

  1. R. J. K. Jacob. What you look at is what you get: eye movement-based interaction techniques. CHI 1990. doi:10.1145/97243.97246: the Midas touch problem.
  2. Juntunen et al. Orchestrating light. Lighting Research & Technology, 2017. doi:10.1177/1477153516654516
  3. Schwarz, Cunningham & Woodford-Smith. Towards a new paradigm for intuitive theatrical lighting control. EVA London 2019. doi:10.14236/ewic/EVA2019.68
  4. Zhang et al. MediaPipe Hands: On-device Real-time Hand Tracking. 2020. arXiv:2006.10214