Loss-matched comparison of strict and adaptive thresholding

Determine the effect of the unlabeled-loss formulation on the comparison between strict fixed-threshold filtering and CW-BASS v2 adaptive thresholding by evaluating a strict fixed-threshold arm within the CW-BASS v2 training loop, thereby isolating the threshold rule from the dual-strong-view and loss-form differences.

Background

The paper compares the UniMatch V2 strict threshold, which uses plain cross-entropy on two independent strong views, with CW-BASS v2 adaptive variants that use one strong view, a feature-perturbation stream, confidence-weighted cross-entropy, and a Sobel-boundary term. Although the authors measure some components of this recipe difference and estimate that they explain only part of the performance gap, they do not ablate the dual-strong-view difference.

A strict-threshold run using the CW-BASS v2 loop would establish whether the observed superiority of strict filtering is attributable to the threshold rule itself or is partly caused by the different unlabeled-loss formulation and augmentation structure. The authors identify this as the first and most valuable remaining experiment.

References

What we cannot bound from our runs is the dual-strong-view difference, which we do not ablate; a drop-in strict arm inside the CW-BASS~v2 loop would settle it and is the first item of future work (Sec.~\ref{sec:limitations}).

— CW-BASS v2: Saturation-Aware Pseudo-Label Selection for Semi-Supervised Segmentation under Foundation-Model Teachers  (2608.12773 - Tarubinga, 13 Aug 2026) in Section 5.3, “What Else Could Explain the Gap,” paragraph “Not (mostly) the loss form or the calibration split”; Section 6, “Scope and limitations”