PixCon: Clean-Positive Contrastive Learning for Foundation-Model Semi-Supervised Segmentation

PixCon is a new semi-supervised semantic segmentation framework that achieves high accuracy by using clean-positive pixel-contrastive learning. It improves over existing methods, including a strong DINOv2-based UniMatch V2 baseline. The framework works well across various datasets, including Pascal VOC, Cityscapes, and ADE20K.

1sec.ai