CW-BASS v2

The one question

Semi-supervised segmentation turns on one question: which pseudo-labels to trust, and how much. A generation of selection rules — dynamic thresholds, per-class curricula, soft confidence weights — answered it for the noisy, under-confident ResNet teachers of the day. Self-supervised foundation encoders (DINOv2) change the regime: the teacher is now strong and its confidence saturates, so filtering that helped a weak teacher can hurt a strong one.

Method

A single DINOv2-Base backbone drives an image-strong and a feature-perturbation stream (UniMatch V2 style) into a DPT-lite decoder. On top, CW-BASS v2 adds three components and one decision:

• Confidence-weighted, boundary-aware self-training — the CW-BASS objective (Sobel boundary auxiliary).
• Held-out calibration slice — 5% of the labeled set, withheld from the supervised gradient, gives an unbiased estimate of pseudo-label noise (in-batch estimates are optimistically biased).
• Self-adaptive confidence floor — with a bounded-retention property, so the dynamic threshold cannot drift to its clamp and admit everything.
• The reliability gate — one forward pass on the held-out slice decides strict vs. floor.

Results

Accuracy tracks the backbone, and CW-BASS v2 sits in the top DINOv2 tier. On the saturated Pascal and Cityscapes teachers the gate selects strict (reproducing the UniMatch V2 operating point); on the confidently-unreliable ADE20K teacher it selects the floor and pulls ahead.

The gate, measured blind

On a held-out slice we measure, per teacher, saturation S = Pr[c ≥ 0.95] and reliability π_kept = Pr[correct | c ≥ 0.95] — a forward-only diagnostic that touches no mIoU. All six DINOv2 teachers are saturated, but they differ in reliability: on Pascal the confident set is trustworthy (π_kept ≈ 98%), so strict filtering is right; on ADE20K it is confidently unreliable (π_kept ≈ 89%, well below the 95% strict assumes), so the floor recovers accuracy a strict cutoff throws away. The gate reads this single statistic and selects the accuracy-optimal rule on all six teachers — with no access to validation mIoU.

Qualitative results

Left: input. Right: CW-BASS v2 prediction overlay (DINOv2-Base, 1/8 split). Validation images, the rule the gate selects on each dataset.

Pretrained models & demo

Each released checkpoint is the EMA teacher of the rule the gate selects on that dataset. Try them in the browser — upload an image, pick a label set:

BibTeX

Copy@article{tarubinga2026cwbassv2,
  title   = {CW-BASS v2: Saturation-Aware Pseudo-Label Selection for
             Semi-Supervised Segmentation under Foundation-Model Teachers},
  author  = {Tarubinga, Ebenezer},
  year    = {2026},
  note    = {Preprint; arXiv ID to be added}
}

v1: CW-BASS, Confidence-Weighted Boundary-Aware Learning for Semi-Supervised Semantic Segmentation, IJCNN 2025 (arXiv:2502.15152).