Testing and failure
Test with images the model has never seen, taken in the conditions where it will really be used. Count how often it is right for each label separately.
Look closely at the mistakes. Do they share something, like dim light or an unusual angle? Those patterns tell you what data to add next.
Decide what happens when the model is unsure. A sorting robot that sets aside low-confidence items for a person to check is safer than one that always guesses.
What should a well-designed system do when the model’s confidence is low?
Choose the answer you think is right.