Define the instance
A bounding-box request needs a rule for what counts as one object. A stack of trays, a person holding a bag, and a product inside transparent packaging can produce different reasonable interpretations. Explain whether nested objects, components, and groups are separate targets. Use annotated examples to show the intended granularity.
Choose the visible boundary
State whether boxes follow visible pixels or an estimated complete object extent. Describe treatment of objects cut off by the image edge and objects hidden behind other objects. Tiny instances need an explicit inclusion threshold tied to the delivered image dimensions. These rules should remain consistent across suppliers and annotation batches.
Preserve coordinate meaning
Specify the coordinate convention, image dimensions, class identifiers, and representation of ignored regions. Resizing or rotating an image changes how coordinates should be interpreted. Request a sample export and render its boxes against the actual delivered files. Successful parsing alone does not show that the labels align with the intended pixels. Include images where no target is present, with an explicit reviewed-negative state so empty annotations cannot be confused with work not yet completed.
Review by failure type
Separate missing objects, incorrect classes, loose boundaries, and duplicate boxes during acceptance review. A single aggregate quality score can conceal an expensive error pattern. Inspect crowded scenes and small objects deliberately. Ask Fiund to review sourcing feasibility against your example set rather than assuming all photographs arrive with suitable detection labels.
Review checklist
- Define object and group boundaries.
- Document visible-versus-complete box policy.
- Render a sample annotation export.
- Review missing objects separately from box precision.