Organize product photos and incoming images by the categories your business uses: product type, visible condition, or a review category. Define the labels, test a model, and inspect the exceptions.
A small catalog team wants to sort incoming images by product type. Start with categories the team already uses, then send uncertain predictions to a person.
Illustrative scenario, not a customer result.A useful classification task connects each output to a next step. An image might be routed to a specialist, assigned a product category, or marked for further inspection. If several categories can apply, clarify that requirement before building a single-label dataset.
Images from one class may share a background, a watermark, or a capture device. A model can learn those shortcuts instead of the visual feature you care about. Review how images were collected and separate evaluation data by source or capture session where appropriate.
A confusion matrix shows which categories are commonly mixed up. Inspect those examples with a domain expert: the labels may overlap, the image may lack enough detail, or the dataset may be missing a relevant condition. Improve the data before assuming a larger model is the answer.
There is no reliable universal minimum. The number depends on the variation in the images, the similarity of the classes, and the cost of mistakes. Start with a representative labeled sample, evaluate a baseline, and use its errors to decide which additional examples to collect.