Takahiro Mano

Mixing labeled data directly into the ClassMix semi-supervised segmentation pipeline yields surprisingly strong gains, suggesting that even imperfectly aligned labeled data is more valuable than noisy pseudo-labels.