CW-BASS v2：基礎模型教師下半監督分割的飽和度感知偽標籤選擇

摘要

半監督語義分割長期以來圍繞著一個核心問題：哪些偽標籤值得信賴。一系列選擇規則——動態閾值、每類別課程式學習、軟置信度權重——在當時解決了雜訊多且置信度不足的 ResNet 教師模型所帶來的挑戰。自監督基礎編碼器改變了這一格局：當使用 DINOv2 教師模型時，置信度趨於飽和，因此曾經有助於弱教師的過濾機制反而可能傷害強教師。我們提出 CW-BASS v2，一種飽和感知的偽標籤選擇方法，它讀取教師模型的置信度狀態，而非固守單一規則。該方法將留出校準（一種無偏的每類別雜訊估計）與可證明能將保留率限制在 1 以下的自適應置信度下限相結合，並在單次通過的門控中加以整合：在留出切片上量測教師模型置信集合的可靠性，即 π_kept = Pr[correct | c ≥ τ]，當其滿足所需置信度（π_kept ≥ τ）時進行嚴格過濾，否則退回到自適應下限。此邊界是既有的操作閾值，而非針對 mIoU 調校的數值；在六個 DINOv2 教師模型上，該方法能在盲測條件下做出正確的嚴格／下限取捨。因此，CW-BASS v2 透過選擇嚴格模式，在飽和的基準數據集上恢復了 UniMatch V2 的操作點（Pascal VOC 1/8 上為 87.4，其報告值為 87.9；Cityscapes 差距在 0.5 以內），並在置信集合不可靠之處（π_kept ≈ 89%，ADE20K）取得改進，此時下限模式略勝一籌（單一種子下 mIoU +1.5）。此門控具有原則性，因為它所避免的失敗是被量測的，而非假設的：在可靠且飽和的教師模型上，置信度分布的動態範圍急劇收縮（Pascal 中 98% 的像素 ≥ 0.95），因此自適應截止值會淹沒保留遮罩，使自訓練退化為確認偏誤。

English

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.