CW-BASS v2: Saturation-Aware Pseudo-Label Selection for Semi-Supervised Segmentation under Foundation-Model Teachers

Hugging Face Daily Papers Papers
semi-supervised-learning semantic-segmentation pseudo-label-selection foundation-models calibration dino-v2

Summary

This paper introduces CW-BASS v2, a saturation-aware pseudo-label selection method for semi-supervised semantic segmentation that adaptively switches between strict filtering and an adaptive confidence floor depending on the teacher's reliability. It shows improved results over baselines across several benchmarks with DINOv2 teachers.

Semi-supervised semantic segmentation has long turned on one question, which pseudo-labels to trust, and a generation of selection rules, dynamic thresholds, per-class curricula, soft confidence weights, answered it for the noisy, under-confident ResNet teachers of their day. Self-supervised foundation encoders change the regime: with a DINOv2 teacher, confidence saturates, so the filtering that helped a weak teacher can hurt a strong one. We propose CW-BASS v2, a saturation-aware pseudo-label selection method that reads the teacher's confidence regime rather than committing to one rule. It pairs held-out calibration, an unbiased per-class noise estimate, with a self-adaptive confidence floor that provably bounds retention away from 1, and combines them in a one-pass gate: measure the reliability of the teacher's confident set, pi_kept = Pr[correct | c >= tau], on a held-out slice, and filter strictly when it meets the confidence demanded (pi_kept >= tau), falling back to the adaptive floor otherwise. The boundary is the pre-existing operating threshold, not a value tuned to mIoU, and across six DINOv2 teachers it makes the correct strict-vs-floor call blind. CW-BASS v2 thus recovers the UniMatch V2 operating point on the saturated benchmarks by selecting strict (Pascal VOC 1/8 87.4 against its reported 87.9; Cityscapes within 0.5), and improves on it where the confident set is unreliable (pi_kept ~ 89%, ADE20K), where the floor edges ahead (+1.5 mIoU, single seed). The gate is principled because the failure it avoids is measured, not assumed: on a reliable, saturated teacher the confidence distribution's dynamic range collapses (98% of Pascal pixels >= 0.95), so an adaptive cutoff floods the retention mask and self-training decays into confirmation bias.
Original Article
View Cached Full Text

Cached at: 08/14/26, 07:27 AM

Paper page - CW-BASS v2: Saturation-Aware Pseudo-Label Selection for Semi-Supervised Segmentation under Foundation-Model Teachers

Source: https://huggingface.co/papers/2608.12773

Abstract

CW-BASS v2 selects pseudo-labels by measuring teacher reliability on held-out data and applying either strict filtering or an adaptive floor to avoid confirmation bias under saturated confidence.

Semi-supervised semantic segmentationhas long turned on one question, which pseudo-labels to trust, and a generation of selection rules, dynamic thresholds, per-class curricula, soft confidence weights, answered it for the noisy, under-confident ResNet teachers of their day.Self-supervised foundation encoderschange the regime: with aDINOv2teacher, confidence saturates, so the filtering that helped a weak teacher can hurt a strong one. We propose CW-BASS v2, a saturation-awarepseudo-label selectionmethod that reads the teacher’s confidence regime rather than committing to one rule. It pairsheld-out calibration, an unbiased per-class noise estimate, with a self-adaptive confidence floorthat provably bounds retention away from 1, and combines them in a one-pass gate: measure the reliability of the teacher’s confident set, pi_kept = Pr[correct | c >= tau], on a held-out slice, and filter strictly when it meets the confidence demanded (pi_kept >= tau), falling back to the adaptive floor otherwise. The boundary is the pre-existing operating threshold, not a value tuned tomIoU, and across sixDINOv2teachers it makes the correct strict-vs-floor call blind. CW-BASS v2 thus recovers the UniMatch V2 operating point on the saturated benchmarks by selecting strict (Pascal VOC 1/8 87.4 against its reported 87.9; Cityscapes within 0.5), and improves on it where the confident set is unreliable (pi_kept ~ 89%, ADE20K), where the floor edges ahead (+1.5mIoU, single seed). The gate is principled because the failure it avoids is measured, not assumed: on a reliable, saturated teacher the confidence distribution’s dynamic range collapses (98% of Pascal pixels >= 0.95), so an adaptive cutoff floods the retention mask and self-training decays intoconfirmation bias.

View arXiv pageView PDFProject pageGitHub0Add to collection

Get this paper in your agent:

hf papers read 2608\.12773

Don’t have the latest CLI?curl \-LsSf https://hf\.co/cli/install\.sh \| bash

Models citing this paper3

#### psychofict/cwbass-v2-pascal Image Segmentation• Updatedabout 3 hours ago #### psychofict/cwbass-v2-cityscapes Image Segmentation• Updatedabout 3 hours ago #### psychofict/cwbass-v2-ade20k Image Segmentation• Updatedabout 3 hours ago

Datasets citing this paper0

No dataset linking this paper

Cite arxiv.org/abs/2608.12773 in a dataset README.md to link it from this page.

Spaces citing this paper0

No Space linking this paper

Cite arxiv.org/abs/2608.12773 in a Space README.md to link it from this page.

Collections including this paper0

No Collection including this paper

Add this paper to acollectionto link it from this page.

Similar Articles

Design-Based Supervised Learning with Noisy Human Labels

arXiv cs.AI

Proposes Partially Adjudicated Design-Based Supervised Learning (PA-DSL), a method that corrects noisy human labels using a small set of adjudicated cases to debias automated classifiers, achieving nominal coverage and reducing RMSE by 10-17% in experiments.