Research map / Label-efficient and robust learning / Semi-supervised learning

Beyond Confidence: Exploiting Homogeneous Pattern for Semi-Supervised Semantic Segmentation

ICML 2025 · Label-efficient and robust learning › Semi-supervised learning · approach: Consistency training for semi-supervised segmentation

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes AgScore, an embedding-space pseudo-label scoring function that uses clean positive and negative agent sets (confidence-selected and orthogonally-selected pixels) and a GMM over the score, integrated into FixMatch, UniMatch and RankMatch for semi-supervised semantic segmentation.

Problem

Confidence-based pseudo-label filtering in semi-supervised semantic segmentation suffers from a trade-off between true positive rate and false positive rate, discarding correct low-confidence labels or admitting incorrect ones.

Approach

Score pseudo labels by their embedding-space similarity to clean positive agents (high-confidence correct pixels) versus negative agents (low-confidence pixels chosen via orthogonal selection), fit a two-component GMM to the resulting score, and use the posterior of being correct as the pseudo-label selection criterion instead of a fixed confidence threshold.

Key results

• On PASCAL classic with ResNet-101 at 1/16 (92 labels), FixMatch improves from 63.9 to 67.0 mIoU with AgScore (+3.1).
• UniMatch with AgScore reaches 75.6 vs 73.5 mIoU at 1/16 on PASCAL classic ResNet-101 (+2.1).
• On Cityscapes ResNet-101 1/16, FixMatch+AgScore improves from 74.2 to 77.6 mIoU (+3.4).

Measured effects reported in the paper

Surprising findings

• Increasing the number of negative agents beyond a point hurts performance because added negatives semantically overlap with positives.
• Increasing positive agents beyond a certain limit degrades performance due to incorrect but high-confidence agents contaminating the positive set.

Limitations

• Gains on stronger baselines such as RankMatch are small (e.g., +0.4 to +1.5 on some PASCAL settings).
• Theoretical analysis relies on distributional approximations (binomial to normal, log-normal ratio) and a discrete proxy of AgScore.

Experimental setup

Datasets: PASCAL VOC 2012 SBD Cityscapes COCO

Benchmarks: PASCAL VOC 2012 classic PASCAL VOC 2012 blender Cityscapes COCO

Metrics: mIoU

Baselines: FixMatch UniMatch RankMatch AugSeg DAW PCR GTA-Seg DDFP PRCL

Builds on: FixMatch UniMatch RankMatch DivideMix

Closest papers by meaning

• Semi-Supervised Semantic Segmentation Using Unreliable Pseudo-Labels (CVPR 2022 · similarity 0.89)
• DAW: Exploring the Better Weighting Function for Semi-supervised Semantic Segmentation (NeurIPS 2023 · similarity 0.88)
• RankMatch: Exploring the Better Consistency Regularization for Semi-supervised Semantic Segmentation (CVPR 2024 · similarity 0.88)
• Improving Semi-Supervised Semantic Segmentation with Sliced-Wasserstein Feature Alignment and Uniformity (CVPR 2025 · similarity 0.87)
• RWMS: Reliable Weighted Multi-Phase for Semi-supervised Segmentation (AAAI 2024 · similarity 0.87)
• SAM as the Guide: Mastering Pseudo-Label Refinement in Semi-Supervised Referring Expression Segmentation (ICML 2024 · similarity 0.87)