Name the required output
Distinguish assigning a category to pixels from separating individual objects. Two adjacent chairs may share a semantic label while requiring different instance identifiers. Describe whether background regions matter and whether every pixel must receive a label. This determines what annotation effort your purchasing brief is actually requesting from a supplier.
Write edge-case rules
Hair, glass, shadows, reflections, and motion blur make boundaries difficult to interpret. Define which visible effects belong to the target region and which should be excluded or flagged. For partially hidden objects, say whether the mask includes only visible pixels. Keep a short reference gallery of accepted and rejected boundaries.
Protect the original geometry
Require masks to match the delivered image dimensions or include a documented mapping. A mask created before cropping can become unusable after an undocumented export. Test holes, disconnected components, and overlapping instances in the intended interchange format. Preserve stable image and instance identifiers so revised annotations can be compared without guessing filenames. Specify whether touching instances share a boundary and how empty masks are represented when an expected target is absent from an otherwise accepted image.
Review expensive regions early
Use the pilot to inspect the regions most likely to consume annotation time. Complex boundaries and crowded images deserve different review attention from isolated objects on plain backgrounds. Agree on acceptable uncertainty and correction procedures before scaling. Commercial availability and the suitability of any proposed masks remain questions for the specific sourcing project.
Review checklist
- Choose semantic, instance, or combined labels.
- Define transparent and ambiguous boundary treatment.
- Verify mask dimensions and instance identifiers.
- Review difficult examples before approving expansion.