Marker detection
The camera finds a known image in the frame and reports its pose. Six frames of stable tracking before we accept it, so a glimpse across the room does not start a lesson.
An instrument that
stays where you put it
The difference between AR that impresses for ten seconds and AR that teaches is whether the thing holds still. Here is the pipeline that makes it hold still.
Plane detection
A downward raycast finds the horizontal surface the marker is lying on. Two and a half seconds is the budget; past that we fall back to the marker pose alone.
World anchoring
The workbench is parented to a world anchor, not to the tracked image. That single decision is why it does not jitter when the marker leaves frame.
Live simulation
192 vertices per trace, recomputed every frame from a pure sampling function, pushed to the GPU in a single call. No allocation, no garbage, no stutter.
Try it
The same maths, in your browser
These are the five sampling functions the Unity build runs, ported verbatim. Drag the sliders and watch the readouts follow.
Live signalSine
Frequency
3.0 Hz
Amplitude
0.70
3.0 Hz
Period 0.33 s · higher frequency packs the crests closer together
0.70
Headroom is fine — clipping starts above 0.94
Waveform character
- Nyquist minimum
- 6.0 Hz
- Third harmonic
- 9.0 Hz
- Level
- −3.1 dB
HologramGlass panel
Readout
3.2 Hz
Readout
0.72
Readouts are projected to screen space each frame, not world-space canvases — cheaper, and legible at every distance.
The workbench
Panels that never hide the instrument
The plot sits on a glass hologram above the marker. Four readouts float around it — frequency, amplitude, period, waveform — positioned so none of them can cover the trace from any viewing angle you are likely to take.
The control dock lives at the bottom of the screen rather than in world space, because a slider you have to reach into the scene for is a slider nobody uses. Drag its handle down and it collapses to a strip, leaving the whole view to the wave.
- Hologram size
- 300 × 128 mm at the anchor plane
- Hover height
- 160 mm above the marker
- Readouts
- Screen-space projection of four 3D anchors
- Dock
- Collapsible, drag threshold 30 pt
Phase 3
A face becomes a signal
The third phase is the one students remember. The AR session switches to the front camera, tracks a face mesh, and reduces it to 68 ordered landmarks. Those landmarks are then unrolled into a single one-dimensional trace.
That is the whole lesson in one gesture: a shape in space becomes a signal in time. Tilt your head and the trace shifts. Move closer and — because the vector is normalised by inter-ocular distance — it does not. Which is exactly the point about why normalisation exists.
No face data leaves the device, and none is stored. The mesh is computed, converted to a feature vector, drawn, and discarded within the same frame. There is no capture step and no upload path in the code.
Unrolled feature trace
Privacy
The camera is a sensor, not a recorder
0
frames written to disk
0
network calls during a session
0
third-party analytics SDKs
Marker detection, plane tracking and the face mesh all run in the device's own AR runtime. The app never gains access to raw frames it could store, and there is no code path that would send one anywhere.
Watch the whole sequence
The prototype runs marker scan, plane lock and workbench anchoring end to end.