Dataset bias in vision

How image/video datasets under-represent people, places, and conditions—and how that shows up as unequal error rates.

Vision deployed on people without bias analysis harms. Measurement is mandatory for serious systems.

Audit who appears in train/test; slice metrics by group/condition (lighting, geography); fix data and objectives—not only marketing claims.

Pick a public vision dataset card; list three representation gaps.

⚠️ Myth: Bigger datasets erase bias.

✓ Reality: Bigger can scale the same skew.

• Course 19 bias-fairness; Course 02 datasheets
• NIST AI RMF: https://www.nist.gov/itl/ai-risk-management-framework ↗