CW-BASS v2: Saturation-Aware Pseudo-Label Selection for Semi-Supervised Segmentation under Foundation-Model Teachers
CW-BASS v2 decides when strong DINOv2 teachers are too confident for the usual pseudo-label filters.
The paper says confidence saturation can make adaptive filtering retain too much, pushing self-training toward confirmation bias. Its gate checks held-out reliability for the teacher’s confident pixels, then chooses strict filtering or an adaptive floor without tuning the boundary to mIoU. Across six DINOv2 teachers, it says the method made the strict-vs-floor call blind. Reported results include near-UniMatch V2 performance on saturated Pascal VOC and Cityscapes settings, and a single-seed +1.5 mIoU gain on ADE20K when the confident set was less reliable. HF Daily Papers' note
The paper says confidence saturation can make adaptive filtering retain too much, pushing self-training toward confirmation bias. Its gate checks held-out reliability for the teacher’s confident pixels, then chooses strict filtering or an adaptive floor without tuning the boundary to mIoU. Across six DINOv2 teachers, it says the method made the strict-vs-floor call blind. Reported results include near-UniMatch V2 performance on saturated Pascal VOC and Cityscapes settings, and a single-seed +1.5 mIoU gain on ADE20K when the confident set was less reliable. HF Daily Papers' note
score 4