Research map / Detection and segmentation / CNN features
AUCSeg: AUC-oriented Pixel-level Long-tail Semantic 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.
Applies pixel-level AUC optimization (square-loss pairwise ranking) to long-tail semantic segmentation on SegNeXt, with a dependency-graph generalization bound and a Tail-class Memory Bank to cut batch-size memory demands.
Problem
Pixel-level long-tail semantic segmentation: tail classes are overlooked; AUC optimization needs at least one sample per class per mini-batch, which pixel-level coupled labels make infeasible and random sampling makes memory-hungry.
Approach
Pixel-level multi-class (one-vs-one) AUC square loss combined with a generalization bound derived via dependency graphs and fractional vertex covers, plus a Tail-class Memory Bank that stores tail-class pixels from past images and pastes them into mini-batches with missing tail classes.
Key results
- AUCSeg reaches 49.20% overall mIoU on ADE20K vs 47.45% for SegNeXt, with tail-class mIoU 45.52% vs 43.28%.
- On Cityscapes AUCSeg achieves 82.71% overall mIoU (tail 81.67%), and on COCO-Stuff 164K 42.73% overall mIoU.
- Applied to other backbones (DeepLabV3+, EMANet, OCRNet, ISANet), AUCSeg improves overall mIoU by roughly 3.5 to 5.5 points; e.g., ISANet 29.53 to 35.07 overall.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use AUCSeg instead of the strongest baseline | SegNeXt · ADE20K val | overall mIoU: +1.05% | Helped |
| Use AUCSeg instead of the strongest baseline | SegNeXt · Cityscapes val | overall mIoU: +0.27% | Helped |
| Use AUCSeg instead of the strongest baseline | SegNeXt · COCO-Stuff 164K val | overall mIoU: +0.31% | Helped |
| Use AUCSeg instead of the strongest baseline | SegNeXt · ADE20K val | tail mIoU: +1.21% | Helped |
| Use AUCSeg instead of the strongest baseline | SegNeXt · Cityscapes val | tail mIoU: +0.75% | Helped |
| Use AUCSeg instead of the strongest baseline | SegNeXt · COCO-Stuff 164K val | tail mIoU: +0.38% | Helped |
| Add AUC optimization (ablation) | SegNeXt · ADE20K val | overall mIoU: 47.45 -> 48.46 (+1.01%) | Helped |
| Add AUC optimization (ablation) | SegNeXt · ADE20K val | tail mIoU: 43.28 -> 44.70 (+1.42%) | Helped |
| Add T-Memory Bank without AUC optimization (ablation) | SegNeXt · ADE20K val | overall mIoU: 47.45 -> 47.86 | No clear effect |
| Add T-Memory Bank to AUC optimization (ablation) | SegNeXt+AUC · ADE20K val | overall mIoU: 48.46 -> 49.20 (+1.75%) | Helped |
| Add T-Memory Bank to AUC optimization (ablation) | SegNeXt+AUC · ADE20K val | tail mIoU: 44.70 -> 45.52 (+2.24%) | Helped |
| Vary Memory Size (S_M) (ablation) | AUCSeg · ADE20K val | mIoU | No clear effect |
Surprising findings
- Using the T-Memory Bank without the AUC loss yields no significant improvement, indicating the AUC loss is necessary for the gain.
Limitations
- Generalization bound depends on the imbalance factor tau and holds under Lipschitz and covering-number assumptions that are mild but not verified empirically
- Gains on head classes are not consistently best; trade-off between head and tail performance
- Hyperparameters (memory size, sample ratio, resize ratio) were tuned on ADE20K and require tuning per dataset
Experimental setup
Datasets: Cityscapes ADE20K COCO-Stuff 164K
Benchmarks: Cityscapes ADE20K COCO-Stuff 164K
Metrics: mIoU
Baselines: DeepLabV3+ EncNet FastFCN EMANet DANet HRNet OCRNet DNLNet PointRend BiSeNetV2 ISANet STDC
Builds on: SegNeXt one-vs-one multi-class AUC Janson decomposition / dependency graph generalization stratified sampling for AUC
Closest papers by meaning
- Towards Robust Pseudo-Label Learning in Semantic Segmentation: An Encoding Perspective (NeurIPS 2025 · similarity 0.85)
- SFC: Shared Feature Calibration in Weakly Supervised Semantic Segmentation (AAAI 2024 · similarity 0.85)
- Diffusion-Driven Two-Stage Active Learning for Low-Budget Semantic Segmentation (NeurIPS 2025 · similarity 0.84)
- RankSEG-RMA: An Efficient Segmentation Algorithm via Reciprocal Moment Approximation (NeurIPS 2025 · similarity 0.84)
- On the Relaxation of Conditional Independence Assumption for Image Segmentation (NeurIPS 2026 · similarity 0.84)
- Pixels to Tokens: Token Space Efficient Active Learning for Low-Budget Semantic Segmentation (NeurIPS 2026 · similarity 0.84)
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.