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

📄 arXiv: 2608.12773v1 📥 PDF

作者: Ebenezer Tarubinga

分类: cs.CV, cs.LG, eess.IV

发布日期: 2026-08-13

备注: Submitted to IEEE TPAMI. 22 pages, 11 figures, 17 tables

💡 一句话要点

提出CW-BASS v2以解决伪标签选择中的饱和问题

🎯 匹配领域: 支柱九：具身大模型 (Embodied Foundation Models)

关键词: 半监督学习 语义分割 伪标签选择 置信度饱和 自适应机制 DINOv2 噪声估计

📋 核心要点

  1. 现有的半监督语义分割方法在伪标签选择上存在挑战，尤其是在强教师模型下，置信度饱和导致选择不准确。
  2. CW-BASS v2通过结合持出校准和自适应置信度下限，提出了一种新的伪标签选择机制，能够动态调整过滤策略。
  3. 在多个数据集上，CW-BASS v2在选择严格伪标签时表现出色，尤其在不可靠的置信集上提升了mIoU性能。

📝 摘要（中文）

半监督语义分割长期以来面临一个核心问题，即如何选择可信的伪标签。传统方法在面对弱教师时使用动态阈值和软置信度权重来过滤噪声标签。然而，随着自监督基础编码器的出现，教师的置信度饱和现象使得这些过滤方法可能对强教师产生负面影响。为此，本文提出CW-BASS v2，一种饱和感知的伪标签选择方法，通过读取教师的置信度状态而非依赖单一规则，结合无偏的每类噪声估计和自适应置信度下限，显著提升了伪标签选择的准确性。

🔬 方法详解

问题定义：本文解决的问题是如何在半监督语义分割中选择可信的伪标签，尤其是在强教师模型下，传统方法的过滤机制可能导致错误选择。

核心思路：CW-BASS v2的核心思路是根据教师的置信度状态动态调整伪标签选择策略，而不是依赖固定的规则。通过结合持出校准和自适应置信度下限，能够有效应对置信度饱和带来的挑战。

技术框架：该方法的整体架构包括两个主要模块：一是持出校准，用于估计每类的噪声水平；二是自适应置信度下限，用于动态调整伪标签的选择标准。通过一遍式的门控机制，评估教师的置信集的可靠性。

关键创新：CW-BASS v2的创新在于其动态调整的伪标签选择机制，能够在置信度饱和的情况下有效避免确认偏差，与传统方法相比，具有更高的灵活性和准确性。

关键设计：该方法的关键设计包括无偏的每类噪声估计和自适应置信度下限的设置，确保在不同置信度条件下都能保持较高的选择准确性。

🖼️ 关键图片

fig_0
fig_1
fig_2

📊 实验亮点

CW-BASS v2在多个数据集上表现出色，特别是在Pascal VOC和Cityscapes上，分别达到了87.4和接近87.9的mIoU，且在ADE20K上提升了1.5 mIoU，显示出其在处理不可靠置信集时的优势。

🎯 应用场景

CW-BASS v2在半监督学习和语义分割领域具有广泛的应用潜力，尤其适用于需要处理大量未标注数据的场景，如自动驾驶、医学影像分析和城市环境理解等。其方法的灵活性和准确性将推动相关领域的研究和应用发展。

📄 摘要（原文）

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.