CW-BASS v2: Saturation-Aware Pseudo-Label Selection for Semi-Supervised Segmentation under Foundation-Model Teachers
Ebenezer Tarubinga
cs.CV, cs.LG, eess.IV
2026-08-13
With a saturated DINOv2 teacher, old dynamic thresholds flood the pseudo-label set. CW-BASS v2 uses held-out calibration to choose strict filtering or a relaxed floor.
Semi-supervised semantic segmentation uses a little labeled data and a lot of unlabeled images. The model assigns pseudo-labels to unlabeled images and trains on them. The perennial question is which pseudo-labels to trust.
A generation of selection rules answered it: dynamic thresholds, per-class curricula, soft confidence weighting. These were built for ResNet teachers that were noisy and under-confident, where noise dominates and filtering low-confidence pixels suppresses it.
That assumption collapses in the foundation-model era. With a DINOv2 teacher, 98% of pixels pile above 0.95 confidence, against only 53% for ResNet-50. Confidence saturates, the dynamic range collapses, and the threshold that once discriminated pixels degenerates toward a constant. Retention drifts to one, residual noise floods training, and self-training decays into confirmation bias. The filtering that helped a weak teacher hurts a strong one.
CW-BASS v2 does not bet on one rule. It reads which confidence regime the teacher is in, then chooses. Three components:
Held-out calibration. A small slice of the labeled set (5%) is held out, and per-class noise is estimated only on data the student never trained on. The point is that in-batch estimation is biased: the student was trained on those pixels, so error measured on them is too optimistic. The paper proves the held-out estimator is unbiased. On the ADE20K teacher, confident-set reliability reads 98.4% in-batch versus 89.3% held-out, a nine-point optimism gap.
Self-adaptive confidence floor. A per-class floor scales with the teacher's mean confidence; the final threshold is the larger of the dynamic threshold and the floor. Theorem 1 pins retention at a fixed quantile, so it cannot drift to one as confidence rises.
One-pass saturation gate. At threshold 0.95, measure the confident set's reliability, the probability that a prediction is correct given confidence above the threshold. If reliability is high enough, filter strictly; otherwise fall back to the floor. The boundary is the system's existing operating threshold, not a value tuned to mIoU.
Across six DINOv2 teachers, the gate calls the regime correctly while blind.
| Dataset | strict mIoU | floor mIoU | pikept | Gate decision |
| Pascal VOC 1/8 | 87.4 | 82.32 | about 98% | strict |
| Cityscapes | near-tie | — | high | strict |
| ADE20K | 49.1 | 50.6 | 89.3% | floor |
On Pascal VOC 1/8, strict filtering reaches 87.4, reproducing the operating point of UniMatch V2's reported 87.9 (the current state of the art). Running the floor unconditionally is the worst variant (82.32 against strict's 87.40), which is exactly why the gate is needed. On ADE20K, where the confident set is unreliable (pikept about 89%), the floor edges ahead by 1.5 mIoU.
This is a measurement insight, not a leaderboard grab. The paper explicitly claims no new peak accuracy; on saturated benchmarks, strict filtering already reproduces UniMatch V2. The real contribution is to articulate that selection rules must change as teachers get stronger, and to provide a diagnostic that tells you which rule applies. Anyone running semi-supervised segmentation with a foundation model should know this trap: do not port the ResNet-era thresholds over unchanged.
The author lists a long set, unusually candid. The calibration slice is only about nine images on Pascal 1/8, too few for a confidence interval; ADE20K results are single-seed; the strict arm uses UniMatch V2's loss while the adaptive arm uses its own scaffold, so it is not a single-factor comparison; the scale-family assumption is idealized, and real confidence distributions change shape during training; CAFS and ENCORE code was not re-run for comparison. The gate itself was validated on the validation split rather than the calibration slice, so it validates the criterion, not the deployed system.