Write the class list first
Identify the subjects that matter and define the boundary between neighboring labels. Include a negative or out-of-scope policy where it is relevant. If a product has several visually similar variants, state whether they share a class or require separate labels.
Describe deployment conditions
List viewpoints, lighting, backgrounds, distances, occlusion, and image quality that the model will encounter. A studio product image and a phone photo of a product in use may answer different needs. Use representative examples to clarify the desired style without assuming the examples themselves are licensable.
Set counts by meaningful group
Specify minimum coverage per class or condition rather than only a total image target. Decide how burst images, near duplicates, and repeated subjects count. If diversity is important, ask for the identifiers or metadata needed to measure it.
Connect labels to the task
Classification, detection, and segmentation require different instructions. Agree on the taxonomy, file format, and review method. A sourcing conversation should also cover the proposed commercial use, source permissions, and exclusions for recognizable people or third-party material.
Review checklist
- Subjects and class taxonomy
- Counts and required visual diversity
- Image formats and minimum dimensions
- Annotation type and review rules
- Permitted uses and exclusions