OT Score: Confidence Metric for Source-Free Domain Adaptation
Researchers introduce the Optimal Transport (OT) score, a confidence metric for source-free unsupervised domain adaptation (SFUDA). Current distributional alignment methods using source class-mean features face computational and theoretical limitations, especially in estimating classification performance without target labels. The OT score, derived from Semi-Discrete Optimal Transport alignment, provides principled uncertainty estimates for target pseudo-labels. It is both intuitively interpretable and theoretically rigorous. Experimental results validate its effectiveness.
Key facts
- OT score is a confidence metric for SFUDA.
- Addresses limitations of current distributional alignment methods.
- Derived from Semi-Discrete Optimal Transport alignment.
- Provides principled uncertainty estimates for target pseudo-labels.
- Theoretically rigorous and intuitively interpretable.
- Experimental results demonstrate effectiveness.
- Focuses on source-free unsupervised domain adaptation.
- Uses source class-mean features.
Entities
—