Image annotation labels what a vision model must learn: classes, boxes, polygons, or tags on product photos, defects, shelves, or scans you own. I write the guideline before a large batch starts, so two people do not invent two meanings for the same class. Image annotation is dataset work. Reading fields off invoices in production is document recognition, not this page.

Volume is a separate line in the estimate. A pilot labels a sample you can review; a full batch follows only after the guideline survives that review. I do not promise a crowd of annotators inside the engineering hours. Exports land next to the images with stable file ids. Near-duplicate photos are handled in deduplication so the same object is not both train and validation.

Acceptance is the class list, the guideline, a double-checked sample, and files a training job can join back to the images. A guideline plus a review sample is often 2–4 weeks. The full batch depends on how many images you actually need.