BIAS-101 - Mitigating Bias in UCF-101

Dataset-level demographic bias auditing and rebalancing pipeline for action recognition.

BIAS-101 investigated demographic bias in UCF-101 and introduced rebalancing and auditing steps for stronger fairness and generalization.

BIAS-101 paper overview

Highlights

  • Audited bias dimensions including gender, age, race, body type, and hair color.
  • Built a CLIP/LLaVA annotation workflow with majority-vote smoothing.
  • Quantified bias with Chi-Square testing and a Dominance Ratio metric.
  • Filtered 85k+ videos via YouTube API and added 2k+ clips for rebalancing.

Resources