CW-BASS v2: Saturation-Aware Pseudo-Label Selection for Semi-Supervised Segmentation

ebenworks · hf · 2026-08-14

CW-BASS v2 introduces a novel pseudo-label selection method for semi-supervised segmentation, specifically in scenarios where foundation models act as teachers.

The core of this approach involves measuring teacher reliability on held-out data to select pseudo-labels. To avoid confirmation bias under saturated confidence, the system applies either strict filtering or an adaptive floor strategy, thereby improving the robustness of segmentation training.

Original post →

More from Research

Research channel →