Start the proposed workflow by checking available images, permitted uses and target tasks. Record sources, conditions and annotation status by batch. Inventory real captures, synthetic examples and missing inputs. Use a representative pilot for screening and annotation. Training and site performance require their own evaluation.
Record screening criteria and outcomes
Check file existence, decoding, source references and candidates for duplicates or low information. Keep rules, thresholds, review decisions and the exclusion list. Mark uncertain samples for review so that inclusion and exclusion are traceable.
Annotate for the task
Use the approved class definitions and rules for the pilot. Check geometry, attributes and images without targets. Resolve disputed examples before expanding the batch. After image transformations, check object locations, visibility and agreement between the image and annotations.
Hand over a recorded version
Deliver the image inventory, annotations, label definitions, screening records and open issues. Read back samples with the target tool and check file links and object information. Keep source and purpose identifiers. Use the agreed independent samples and real-data scope for later evaluation.
Related service: Data, training and model engineering.