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

PixCon proposes a clean-positive pixel-contrastive framework for foundation-model semi-supervised semantic segmentation. Instead of filling class memory banks with confidence-filtered pseudo-labels, it admits only labeled pixels that the student currently classifies correctly, giving a construction-level guarantee of zero positive-set contamination, denoted ρF=0. The method adds no inference-time parameters and uses a single branch on a consistency backbone. Experiments on Pascal VOC, Cityscapes, and ADE20K compare PixCon with a DINOv2-based UniMatch V2 baseline, while a first-order InfoNCE analysis quantifies the effect of false positives.