Drop-in evaluation of published CAFS and ENCORE systems

Evaluate the released implementations of the published CAFS and ENCORE per-class thresholding systems, rather than only a faithful reimplementation of their mechanisms, to determine whether the reported negative results generalize to the original systems under foundation-model teachers.

Background

The paper audits the held-out calibration mechanism associated with CAFS and the feedback-driven mechanism associated with ENCORE through its own implementation. However, the experiments do not use the original released code of either published system.

A direct drop-in evaluation of the released systems would test whether implementation-specific choices, undocumented details, or differences in training configuration account for the observed failure of per-class adaptive thresholding under DINOv2 teachers.

References

Finally, we audit the CAFS/ENCORE mechanism (held-out and feedback-driven per-class thresholds) through our own faithful reimplementation (Sec.~\ref{sec:perclass}) rather than their released code; a drop-in run of the published systems is future work.

— 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, “Scope and limitations,” item (v)