Choose representative review images
Include difficult conditions such as small objects, overlap, blur, reflections, and partial visibility. An easy sample may say little about the edge cases in the complete collection. Keep a record of the conditions represented in your review.
Check the instructions behind the labels
Ask how ambiguous examples are handled. For bounding boxes, decide how occluded and truncated objects are treated. For masks, define boundaries and holes. For classification, explain when an image can carry several labels or should be marked uncertain.
Separate label issues from coverage issues
A collection can have carefully drawn annotations and still miss a needed class or environment. Review label consistency and coverage as separate questions. Compare counts with the proposed taxonomy and inspect unannotated examples if the task requires exhaustive labels.
Agree on correction handling
Define how findings are reported and what a corrected version contains. Stable image and annotation identifiers help preserve the connection between review comments and delivered files. Document taxonomy changes so an updated label set does not silently alter the meaning of previous work.
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