Project: train an image classifier in the browser

The project: build an image classifier in the browser (Teachable Machine) with three classes of your own choosing — scissors/pencil/eraser, say, or three hand signs.

Requirements:

1. At least 30 pictures per class, with variation.
2. Set aside 10 pictures per class that the model may not train on (the test set).
3. Measure the accuracy on the test set: correct / 30.
4. Find at least three pictures the model classifies wrongly and explain why.

A model without a test set is a guess about how good it is.

A reporting template:

Classes	scissors, pencil, eraser
Training pictures	35 / 32 / 40
Test pictures	10 / 10 / 10
Accuracy	26/30 = 87 %
Mistakes	pencil→eraser ×2 (a short pencil), scissors→pencil (closed scissors) …
Improvement	more pictures of closed scissors; test again

Also write: what would happen if someone else used the model with their objects and their lighting?