Teach it two things
Show it a few examples of each. Then it guesses, live.
Google's Teachable Machine (2017) let anyone teach a computer with a webcam and no code. I found the idea in December 2025 and rebuilt it here on SigLIP 2, a newer image model.
The model turns each picture into 768 numbers. Your examples stay in this page as those numbers. Each new frame is compared with them, after taking away what they all share (the room, the light), and goes to the class it sits closest to. Nothing is trained or uploaded; a few examples are enough.
The practice camera draws a pretend scene, for when you have no camera or would rather not. Photos you drop on a class teach it too; drop one on the picture to test it.
Runs on your device. Your camera and photos never leave it.
Show it a few examples of each. Then it guesses, live.