For vision annotation procurement, define the task and unit of annotation. Classification, bounding boxes, segmentation regions and text annotations each need output rules. Define classes, easily confused objects, areas to annotate and areas to ignore. Include positive, negative and boundary examples.

Test the required fields in a pilot

Specify image identifiers, class identifiers, geometry, attributes, occlusion status and handling of images with no target. Begin with a small, representative annotation batch. Check whether the rules cover the examples. Keep reasons and decisions for disputed objects, then issue the agreed rule version.

Define sampling and rework

State the delivery batch, classes and difficult conditions from which samples are drawn. Check missing objects, wrong classes, geometric deviations and missing attributes using agreed criteria. List classes or conditions absent from the sample. After rework, check the affected scope against the revised delivery.

Hand over the sample inventory, label definitions, annotation files, rule version, review records and open issues. Base acceptance on recorded checks for that batch. Keep separate evidence for counts and format checks.

Related service: Data, training and model engineering.