Research map / Label-efficient and robust learning / Recognition and long-tail
That Label's got Style: Handling Label Style Bias for Uncertain Image Segmentation
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Condition on label style while training with all annotation styles | Prob. U-net · ISIC styles 0, 1, 2 | IoU: 0.65 (0.09), 0.65 (0.10), 0.69 (0.21) -> 0.77 (0.15), 0.78 (0.15), 0.76 (0.15) | Helped |
| Condition on label style while training with all annotation styles | Prob. U-net · PhC-U373 styles 0, 1 | IoU: 0.89 (0.01), 0.90 (0.02) -> 0.92 (0.02), 0.92 (0.01) | Helped |
| Condition on label style while training with all annotation styles | SSN · ISIC styles 0, 2 | IoU: 0.72 (0.01), 0.71 (0.19) -> 0.77 (0.10), 0.78 (0.19) | Helped |
| Condition on label style while training with all annotation styles | SSN · ISIC style 1 | IoU: 0.77 (0.09) -> 0.71 (0.20) | Hurt |
| Condition on label style while training with all annotation styles | SSN · PhC-U373 styles 0, 1 | IoU: 0.89 (0.02), 0.93 (0.02) -> 0.92 (0.01), 0.93 (0.01) | Mixed |
| Condition on label style and train with all styles | Prob. U-net · ISIC label style 0 ground truth | Mean area difference (pixels): 6019 (13648) -> 311 (8332) | Helped |
| Condition on label style and train with all styles | Prob. U-net · PhC-U373 label style 0 ground truth | Mean area difference (pixels): 370 (402) -> 43 (417) | Helped |
| Condition on label style and train with all styles | SSN · ISIC label style 0 ground truth | Mean area difference (pixels): 5176 (15629) -> 3199 (12987) | Helped |
| Condition on label style and train with all styles | SSN · PhC-U373 label style 0 ground truth | Mean area difference (pixels): 716 (1022) -> 262 (558) | Helped |
| Replace coarse annotations with augmented versions of fine-grained annotations (ablation) | Conditioned models · ISIC styles 0, 1, 2 | IoU: c-prob. U-net: 0.77 (0.15), 0.78 (0.15), 0.76 (0.15) vs c-prob. U-net (aug): 0.73 (0.17), 0.70 (0.09), 0.69 (0.14); c-SS | Hurt |
| Replace coarse annotations with augmented versions of fine-grained annotations (ablation) | Conditioned models · ISIC annotator distributions | GED: c-prob. U-net: 0.57 (0.13), 0.55 (0.24), 0.49 (0.19), 0.55 (0.18); c-prob. U-net (aug): 0.59 (0.12), 0.57 (0.21), 0.53 ( | Hurt |
| Use only style 0 annotations instead of all styles (ablation) | Prob. U-net and SSN · ISIC and PhC-U373 | Area bias | Mixed |
Closest papers by meaning
- Towards Fairness under Label Bias in Image Segmentation: Impact, Measurement and Mitigation (NeurIPS 2026 · similarity 0.90)
- ValUES: A Framework for Systematic Validation of Uncertainty Estimation in Semantic Segmentation (ICLR 2024 · similarity 0.90)
- Uncertainty-aware Fine-tuning of Segmentation Foundation Models (NeurIPS 2024 · similarity 0.90)
- UNICL-SAM: Uncertainty-Driven In-Context Segmentation with Part Prototype Discovery (CVPR 2025 · similarity 0.87)
- UncertainSAM: Fast and Efficient Uncertainty Quantification of the Segment Anything Model (ICML 2025 · similarity 0.87)
- Flaws can be Applause: Unleashing Potential of Segmenting Ambiguous Objects in SAM (NeurIPS 2024 · similarity 0.86)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.