Licensing checklists

Inspection image licensing: labels, duplicates, and splits

Prepare inspection images with part-level counts, capture context, label provenance, and evaluation splits that expose coverage gaps.

Image counts can hide near-duplicates and repeated views of the same part. Explain the inspection setup, how images relate to physical items, and how the archive was grouped. The goal here is to define a reviewable collection and a delivery plan before discussing commercial terms.

Identify the physical unit behind each image

Record image ID, part or batch ID, capture time, camera or station, and view when available. Separate image counts from distinct physical-part counts. Multiple near-identical frames can inflate the apparent size of a collection. Report missing parent IDs instead of inferring relationships that the records cannot support.

Describe how labels were obtained

Keep the class taxonomy and its version with the collection. Distinguish labels observed during inspection, labels assigned later, and labels inferred from another system. Record reviewer roles, disagreement handling, and unknown or ambiguous cases. Do not present an unreviewed machine-generated label as a confirmed physical defect.

Measure coverage at the level of the task

Show counts for normal examples, each defect class, capture station, lighting condition, and time period where known. Indicate whether a count refers to images or parts. Rare-defect enrichment can help test a particular failure mode, but it does not reproduce the operating prevalence of that defect. State that limitation in the sample notes.

Split related images together

Ask the evaluator which grouping avoids overlap for its intended task. Keeping all views of a physical part together is a useful starting point; batch or time grouping may also matter. Record the grouping rule and unresolved identity gaps. Compare both metadata and obvious duplicates before claiming that evaluation subsets are independent.

Review visible context and deliver label geometry

Flag faces, badges, serial numbers, customer identifiers, and restricted facility context for appropriate review. Document bounding-box coordinate conventions or mask formats, image orientation, and any resizing applied. Check that annotations still align after transformations and keep transformed files linked to the original version.

Inspection image review sheet

Edit this outline, then copy or download it. It stays in this page and is not submitted or saved by this tool. Keep confidential records and personal information out. Save your work before leaving or reloading.

You can select and copy the text manually. This is a preparation worksheet, not a legal agreement.

Sources and scope of this guidance

Sources checked October 19, 2026. AWS describes field dictionaries as an aid to evaluating a dataset. The workflow and examples here are our preparation suggestions, not AWS requirements or a claim of platform eligibility.

AWS Data Exchange: data dictionaries