Describe the expected distribution
Equal class counts may be useful for a controlled experiment, but they are not automatically representative of a deployed workload. Describe the expected frequency of categories and the cost of missing each one. Keep acquisition goals separate from evaluation sampling so you can inspect rare classes without confusing that sample with real-world prevalence.
Measure diversity inside the count
A class with many images can still contain little meaningful variation. Ask how examples differ by object instance, setting, capture session, and visible condition. For a rare category, related shots may be easier to find than independent examples. Make the unit of coverage explicit before treating a numeric target as satisfied by a supplier.
Handle scarcity transparently
Identify which categories are mandatory and which can be merged, deferred, or collected separately. Do not silently broaden a label to hit a quantity target. Review sparse categories with concrete examples and document what remains missing. Feasibility discussions are more productive when the brief allows tradeoffs without obscuring the original requirement or label meaning. Revisit priorities when the task changes instead of preserving an obsolete target merely because it was convenient in the first acquisition round.
Review performance by slice
Plan an evaluation report that separates categories and important conditions. Aggregate results can look acceptable while a small class remains untested. Use pilot findings to decide whether the next acquisition should add volume, distinct subjects, or a specific difficult condition. Fiund sourcing conversations should evaluate those gaps rather than assume an advertised image count resolves them.
Review checklist
- Separate acquisition targets from deployment prevalence.
- Count independent examples within each class.
- Document scarcity and acceptable compromises.
- Evaluate important classes and conditions separately.