PixCon

Abstract

Semi-supervised semantic segmentation has long turned on one question — which pseudo-labels to trust — and answered it with ever more careful confidence filtering. Foundation backbones change the regime: with a DINOv2 teacher a strict threshold already retains a measured 98%-clean pseudo-label set, so the accuracy that remains lives not in the filter but in how the embedding space is structured by class.

We propose PixCon, a clean-positive pixel-contrastive framework. PixCon maintains a per-class memory bank that admits only labeled pixels the student already classifies correctly, guaranteeing a contamination-free positive set (ρ_F = 0) by construction — unlike prior contrastive banks (ReCo, U²PL) built from confidence-filtered pseudo-labels. It is a single branch over a consistency backbone, adds no inference-time parameters, and needs no bank-specific threshold.

Purity over abundance

A first-order analysis of the supervised-InfoNCE gradient shows the false-positive term scales as ρ_F / (1 − ρ_F), where ρ_F is the fraction of contaminated positives. Rather than assume it, we measure it. The admission rule is deliberately strict:

Because every stored positive is verifiably correct, ρ_F = 0 by construction. Since contamination is already rare under a foundation-model teacher, that guarantee hardens the method as teachers weaken, while cleaner positive supervision drives the accuracy.

Top of the SSSS leaderboard

Semi-supervised mIoU (%), 1/8 split. PixCon sits with the strongest published results and clears the pre-foundation-model field by a wide margin on all three benchmarks.

BibTeX

@article{tarubinga2026pixcon,
  title   = {PixCon: Clean-Positive Contrastive Learning for
             Foundation-Model Semi-Supervised Segmentation},
  author  = {Tarubinga, Ebenezer},
  journal = {arXiv preprint arXiv:2607.03068},
  year    = {2026}
}