CYRA Lab · Module 05
Teach CYRA
Show a sign, and it learns it in seconds.
Here you train the model: show CYRA your own hand sign, let it learn in a few seconds and recognise it right away. There are three sign slots; name them yourself, teach the signs one by one, then put CYRA to the test.
- Runs on your device
- ~11 MB on first start
- Model: MediaPipe Gesture Recognizer · hand points + nearest neighbours (k-NN)
How to try it
Turn on the camera. The panel alongside has three sign slots; rename them if you like.
Press “Teach” and hold your sign for 2–3 seconds; turn your hand slightly so CYRA learns it from different angles. Repeat for each slot.
Now show the signs one by one: CYRA says the name of the sign it recognises and the panel shows how sure it is.
Your camera permission is used only on this page and only when you turn it on.
Your image never leaves your device: no frame is recorded or sent.
Taught signs
- 0 examples
- 0 examples
- 0 examples
Turn on the camera first.
CYRA log
- Once you turn on the camera, we start talking here.
Behind the scenes
- Every frame, the hand model's 21 joint points are aligned to the wrist and divided by the size of the palm, so the same sign gives the same numbers whether your hand is near or far.
- Recognition uses the nearest-neighbours method (k-NN) on the examples you teach: training takes seconds, all in your browser. Only hand points are kept, not images, and everything is deleted when the camera turns off.
- A hand shape you have not taught does not fall into a random name: if even the nearest example is far, CYRA says it does not recognise it. This method is for still, one-hand shapes; it can be a basis for a sign-language vocabulary, which involves movement and both hands, but it is not enough on its own.
Where it helps
- Personal controls: panels and remotes where each user can define their own hand signs.
- Quick setup in the field: teach a device a new movement with a few examples and use it right away.
- Accessibility: a user who cannot speak or hear teaching the device their own signs.
Example use ideas; we design the right one for your project together.
Next module
Show & Tell
Whatever is in the frame, CYRA counts it.
What would you like your project to sense?
Motion, objects, people or surroundings: tell us what you need and let's design the embedded solution together.
Models: Google MediaPipe Gesture Recognizer, Face Landmarker and EfficientDet-Lite0 (COCO dataset), Apache-2.0 licensed. Spoken replies use your browser's own voice.