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

Hugging Face Daily Papers Papers
semi-supervised-learning semantic-segmentation contrastive-learning foundation-models pixel-contrastive memory-banks

Summary

PixCon proposes a clean-positive pixel-contrastive framework for semi-supervised semantic segmentation that guarantees contamination-free positive sets via per-class memory banks, improving accuracy over existing methods on benchmarks like Pascal VOC, Cityscapes, and ADE20K.

Semi-supervised semantic segmentation (SSSS) 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 SSSS banks (ReCo, U^2PL) 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. A first-order analysis of the supervised-InfoNCE gradient explains why contamination hurts: its false-positive term scales as ρ_F/(1-ρ_F), which we measure (0.018 on Pascal, 0.106 on ADE20K) rather than assume. Across Pascal VOC, Cityscapes, and ADE20K, PixCon matches or improves a strong DINOv2-based UniMatch V2 baseline in a compute-matched one-switch protocol: it improves every Pascal-1/8 seed (a per-seed gain of about +0.2 mIoU) and its three-seed mean reaches 87.90, the published UniMatch V2-B figure. Because contamination is already rare under foundation-model teachers, our analysis indicates the ρ_F=0 guarantee acts chiefly as robustness as teachers weaken, while the accuracy gain comes from cleaner positive supervision, making clean-positive contrast a robust, low-cost default for foundation-model SSSS.
Original Article
View Cached Full Text

Cached at: 07/07/26, 06:42 AM

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

Source: https://huggingface.co/papers/2607.03068

Abstract

PixCon is a semi-supervised semantic segmentation framework that uses clean-positive pixel-contrastive learning with per-class memory banks to improve accuracy over existing methods.

Semi-supervised semantic segmentation(SSSS) has long turned on one question, whichpseudo-labelsto trust, and answered it with ever more careful confidence filtering. Foundation backbones change the regime: with aDINOv2 teachera 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-positivepixel-contrastive framework. PixCon maintains a per-classmemory bankthat admits only labeled pixels the student already classifies correctly, guaranteeing acontamination-free positive set (ρ_F=0) by construction, unlike prior contrastive SSSS banks (ReCo, U^2PL) built from confidence-filteredpseudo-labels. It is a single branch over aconsistency backbone, adds no inference-time parameters, and needs no bank-specific threshold. A first-order analysis of thesupervised-InfoNCE gradientexplains whycontaminationhurts: its false-positive term scales as ρ_F/(1-ρ_F), which we measure (0.018 on Pascal, 0.106 on ADE20K) rather than assume. Across Pascal VOC, Cityscapes, and ADE20K, PixCon matches or improves a strong DINOv2-based UniMatch V2 baseline in a compute-matched one-switch protocol: it improves every Pascal-1/8 seed (a per-seed gain of about +0.2 mIoU) and its three-seed mean reaches 87.90, the published UniMatch V2-B figure. Becausecontaminationis already rare under foundation-model teachers, our analysis indicates the ρ_F=0 guarantee acts chiefly as robustness as teachers weaken, while the accuracy gain comes from cleaner positive supervision, making clean-positive contrast a robust, low-cost default for foundation-model SSSS.

View arXiv pageView PDFProject pageGitHub0Add to collection

Get this paper in your agent:

hf papers read 2607\.03068

Don’t have the latest CLI?curl \-LsSf https://hf\.co/cli/install\.sh \| bash

Models citing this paper3

#### psychofict/pixcon-pascal Image Segmentation• Updated31 minutes ago #### psychofict/pixcon-cityscapes Image Segmentation• Updated31 minutes ago #### psychofict/pixcon-ade20k Image Segmentation• Updated31 minutes ago

Datasets citing this paper0

No dataset linking this paper

Cite arxiv.org/abs/2607.03068 in a dataset README.md to link it from this page.

Spaces citing this paper1

Collections including this paper0

No Collection including this paper

Add this paper to acollectionto link it from this page.

Similar Articles

@lukaskuhn77: We introduce LeVLJEPA: the first fully non-contrastive end-to-end vision-language pretraining method competitive with C…

X AI KOLs Following

LeVLJEPA is the first fully non-contrastive end-to-end vision-language pretraining method, achieving competitive performance with CLIP and SigLIP without negatives, temperature, or momentum encoder. It learns via cross-modal prediction with stop-gradient targets and per-modality distributional regularization, providing stronger dense semantic features for downstream tasks like VLM backbones and semantic segmentation.

Vision Pretraining for Dense Spatial Perception

Hugging Face Daily Papers

This paper introduces masked boundary modeling, a self-supervised paradigm for vision pretraining that learns sub-pixel boundary representations to improve dense spatial perception. The resulting model, LingBot-Vision, demonstrates significant improvements in depth estimation and other downstream tasks, showing that boundary modeling is a scalable pretraining principle for spatially structured visual representations.

Representations Before Pixels: Semantics-Guided Hierarchical Video Prediction

Hugging Face Daily Papers

Re2Pix is a hierarchical video prediction framework that improves future video generation by first predicting semantic representations using frozen vision foundation models, then conditioning a latent diffusion model on these predictions to generate photorealistic frames. The approach addresses train-test mismatches through nested dropout and mixed supervision strategies, achieving improved temporal semantic consistency and perceptual quality on autonomous driving benchmarks.