Semi-supervised semantic segmentation has long relied on pseudo-label selection rules—dynamic thresholds, per-class curricula, and soft confidence weights—designed for noisy, under-confident ResNet teachers. However, self-supervised foundation encoders like DINOv2 change the regime: confidence saturates, with 98% of Pascal VOC pixels at or above 0.95 confidence. This saturation can cause adaptive cutoffs to flood the retention mask, degrading self-training into confirmation bias.
CW-BASS v2, introduced in a new paper, addresses this by measuring the teacher's confidence regime rather than assuming one. It uses a one-pass reliability gate that estimates pi_kept = Pr[correct | c >= tau] on a held-out slice. If the confident set is as reliable as the demanded confidence (pi_kept >= tau), it filters strictly; otherwise, it falls back to a self-adaptive floor. The boundary is the pre-existing operating threshold, not tuned to mIoU.
Across six DINOv2 teachers, CW-BASS v2 makes the correct strict-vs-floor call blindly. It recovers the UniMatch V2 operating point on saturated benchmarks: Pascal VOC 1/8 achieves 87.4 mIoU (vs. 87.9 reported), and Cityscapes is within 0.5. On ADE20K, where the confident set is unreliable (pi_kept ~89% vs. 98%), the adaptive floor edges ahead by +1.5 mIoU (single seed).
The method's principle is that the failure it avoids is measured, not assumed. Code, weights for Pascal/Cityscapes/ADE20K, and a Gradio demo are available in the repository.