Research map / Label-efficient and robust learning / Recognition and long-tail

Towards Fairness under Label Bias in Image Segmentation: Impact, Measurement and Mitigation

NeurIPS 2026 · Label-efficient and robust learning › Recognition and long-tail · approach: Group-robust classification · 0 citations (OpenAlex)

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.

Adapts Confident Learning to segmentation to audit group-conditional label bias without clean labels, shows label bias amplifies subgroup separability in encoder features, and mitigates it with subgroup-conditioned FiLM decoding plus asymmetric boundary masking.

Problem

Label bias (group-conditional, systematic annotation errors) in image segmentation causes subgroup performance disparities, but detecting it normally requires clean unbiased labels, which are typically unavailable; standard overlap metrics like Dice/IoU also mask it.

Approach

Projects Confident Learning joint distributions onto subgroup-level and directional metrics (LER, RR, DR) for segmentation to audit label bias from observed labels alone, and uses FiLM subgroup conditioning with boundary masking on the biased group, inferring all outputs in a selected reference annotation style.

Key results

• On CelebAMask-HQ at beta=100%, SCD raises true IoU to 89.0 with subgroup gap 8.9 versus ERM 82.5 with gap 21.0.
• Label bias on CelebAMask-HQ raises bottleneck MMD from 0.10 to 0.29 and centroid distance from 7.54 to 19.04 as beta goes 0% to 100%.
• Observed IoU gap on CelebAMask-HQ at beta=75% shrinks to 10.0% while true gap grows from 5.5% to 17.0%, showing standard metrics mask bias.

Surprising findings

• Standard IoU/Dice against corrupted labels can suggest improved fairness while true disparity increases (biased ruler effect).
• Fairness regularization and feature alignment fail to close the gap and can cause leveling down.

Limitations

• CL audit disagreement may reflect model error or miscalibration rather than true label error, and underestimates severe bias at beta=100%.
• Mitigation assumes a subgroup with a cleaner annotation style can be identified; auto-selection by training loss may reflect task difficulty.
• Label bias simulation covers only a subset of realistic annotation error types; ISIC lacks clean masks for validation.

Experimental setup

Datasets: CelebAMask-HQ PhC-U373 IMA++ ISIC 2017

Benchmarks: CelebAMask-HQ PhC-U373 IMA++

Metrics: Dice IoU subgroup IoU gap Confident Label Error Rate (LER) relative risk (RR) directional rate (DR) silhouette score Fisher ratio MMD linear probe accuracy

Baselines: ERM EO DP Adversarial debiasing MMD alignment Deep CORAL GCE Bootstrapping Asymmetric masking

Builds on: Confident Learning FiLM Gradient reversal (DANN) Generalized Cross Entropy Bootstrapping loss

Closest papers by meaning

• That Label's got Style: Handling Label Style Bias for Uncertain Image Segmentation (ICLR 2023 · similarity 0.90)
• Stratify or Die: Rethinking Data Splits in Image Segmentation (NeurIPS 2025 · similarity 0.88)
• Classifier-to-Bias: Toward Unsupervised Automatic Bias Detection for Visual Classifiers (CVPR 2025 · similarity 0.87)
• Discover and Mitigate Multiple Biased Subgroups in Image Classifiers (CVPR 2024 · similarity 0.86)
• Subgroups Matter for Robust Bias Mitigation (ICML 2025 · similarity 0.86)
• Mitigating Label Bias in Machine Learning: Fairness through Confident Learning (AAAI 2024 · similarity 0.85)