Empirical trajectory of the held-out per-class noise estimate

Characterize the training trajectory of the held-out per-class pseudo-label noise estimates for the classes whose adaptive thresholds are lowered, in order to determine directly whether their estimated noise rates remain stable or increase rather than fall during training.

Background

CW-BASS v2 uses held-out calibration to estimate per-class pseudo-label noise without the downward bias induced by evaluating on training pixels. The paper predicts that the estimated noise rate should not decrease for classes whose adaptive thresholds are lowered, because the lower-confidence pixels admitted by those thresholds are disproportionately erroneous.

The authors rely on per-class IoU evidence as an indirect check and explicitly state that they do not report the direct trajectory of the estimated per-class noise rates. A direct analysis would test the diagnostic prediction more directly and clarify how the noise signal evolves during adaptive self-training.

References

We lean on the measured per-class IoU rather than a direct read of $\widehat\varepsilon_k$, whose trajectory we leave to future work (Sec.~\ref{sec:analysis}).

— CW-BASS v2: Saturation-Aware Pseudo-Label Selection for Semi-Supervised Segmentation under Foundation-Model Teachers  (2608.12773 - Tarubinga, 13 Aug 2026) in Section 6.1, “Anatomy of the Collapse,” paragraph “Link 3: mask flooding, and the admitted pixels are disproportionately wrong”