Head rig
A head whose angle we already know
Every head-pose figure in this prototype was measured against a person doing their best to follow an instruction, with no record of what they actually did. This page removes the person. One photograph becomes a rotatable 3D head, the rendered frames are fed to the same detector an exam uses, and the angle it reports is compared with the angle that was drawn — which is known exactly, because it was typed in.
What this proves, and what it does not
- It proves the geometry and the code. That a known rotation comes back at the right size and the right sign, that eye-only movement leaves head pitch alone, that the extrapolation guard fires where it should, and that a head turning while the eyes stay on the screen is not reported as looking away.
- It does not prove accuracy on real faces. This is a face shell with no hair, ears or neck, lit flat. A detector's behaviour on it is not its behaviour on a person — the same domain gap that limited every synthetic eye dataset. No number from this page belongs in a claim about how well the product works in a room.
- Nothing here is a separate implementation. The rendered canvas is handed to the production monitor as a MediaStream, so the right-hand column below went through the same detector, geometry and classifier as a candidate's exam would.
The rig
Starting…
Click anywhere in the box below to point the eyes at that spot. Outside the box is off the screen, which is the case the whole product exists to detect.
A calibration performed exactly
The rig can aim its eyes at a named point and hold a named head pose, both exactly, so it can complete the nine-point calibration the way no person can. That matters for one question above all: a head-pose fallback was tried on real data once and rejected, because on 334 frames where the candidate was looking at the screen with their head turned it fired on 319 of them. With a model fitted here, "head turns, eyes stay on the screen" stops being an anecdote and becomes a count against a truth that was drawn rather than reported.
Whatever error this reports is a FLOOR, not an estimate. Nothing here blinks, drifts, squints or gets bored, and the face is a flat-lit shell with no hair. A real candidate cannot do better than this number and will do worse.
Scripted cases
Each of these is a sequence of poses with an exact ground truth. Two of them exist because of measured failures rather than imagination: a head-pose fallback was built, measured and rejected once for firing on 319 of 334 frames where the candidate was looking at the screen with their head turned, and the first corner-referenced vertical gaze channel sat inert for two rounds because nothing tested that eye-only movement leaves head pitch alone.
Build the head from your own face
The bundled source frame is 480×360, which is a poor texture. Capturing your own face at 1080p gives a far better one — and watching your own head turn on command is the fastest way to see what the rig is.