Research map / Label-efficient and robust learning / OOD and anomaly detection
EAT: Towards Long-Tailed Out-of-Distribution Detection
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 EAT, a long-tailed OOD detection method that trains outliers with dynamic virtual abstention labels, applies tail-class CutMix onto OOD backgrounds, fine-tunes the classifier with logit-adjusted loss, and ensembles three heads.
Problem
Out-of-distribution detection when the in-distribution training data follows a long-tailed class distribution, where OOD samples are easily confused with tail-class samples.
Approach
Replaces uniform-target outlier exposure with multiple abstention classes assigned to OOD samples via self-generated virtual labels, augments tail classes by pasting tail objects onto context-rich head-class or OOD backgrounds with CutMix, and fine-tunes only the classifier with a logit-adjusted loss.
Key results
- On CIFAR10-LT, averaged over six OOD test sets, EAT reaches 92.87 AUROC and 28.83 FPR95 versus PASCL's 90.99 and 33.36.
- On CIFAR100-LT, EAT reaches 75.45 AUROC and 46.23 ACC95 versus PASCL's 73.32 AUROC and 43.10 ACC95.
- On ImageNet-LT, EAT improves over OE by 3.51% AUROC and 9.19% in-distribution accuracy, and as a plug-in it raises RIDE's AUROC from 66.18 to 72.41.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use EAT instead of PASCL | CIFAR10-LT · CIFAR10-LT average OOD test sets | AUROC: 90.99 -> 92.87 | Helped |
| Use EAT instead of PASCL | CIFAR100-LT · CIFAR100-LT average OOD test sets | AUROC: 73.32 -> 75.45 | Helped |
| Add EAT to the base method | RIDE · CIFAR-100, six OOD datasets | AUROC: 66.18 -> 72.41 | Helped |
| Add EAT to the base method | GLMC · CIFAR-100, six OOD datasets | AUROC: 65.01 -> 73.07 | Helped |
| Add virtual labels for OOD training samples (ablation) | EAT · CIFAR10-LT, SVHN OOD | AUROC: 84.91 -> 97.92 | Helped |
| Add classifier fine-tuning (ablation) | EAT · CIFAR100-LT, SVHN OOD | ACC@FPR0: 43.34 -> 46.23 | Helped |
| Add CutMix tail augmentation (ablation) | EAT · CIFAR100-LT, SVHN OOD | AUROC: 79.52 -> 83.11 | Helped |
| Increase ensemble size from 1 to 3 (ablation) | EAT · CIFAR10-LT, SVHN OOD | AUROC: 97.08 -> 97.92 | Helped |
| Increase ensemble size from 1 to 3 (ablation) | EAT · CIFAR100-LT, SVHN OOD | AUROC: 80.70 -> 83.11 | Helped |
Surprising findings
- A strong long-tailed in-distribution classifier does not imply good OOD detection; RIDE and GLMC lag far behind the proposed method on OOD detection.
- Without virtual labels, some variants achieve higher ACC95, because more in-distribution samples are flagged as OOD, so ACC95 alone is misleading.
Limitations
- CutMix generates additional data, increasing GPU memory and time overhead.
- Tail-class CutMix assigns the foreground label to mixed images, introducing label noise that is mitigated only by lower sample weights.
- Only plain CutMix is used; sampling distributions and other components are not adapted for long-tailed OOD.
Experimental setup
Datasets: CIFAR10-LT CIFAR100-LT ImageNet-LT 80 Million Tiny Images (300K subset) ImageNet-Extra ImageNet-22k
Benchmarks: CIFAR10-LT OOD tests (Textures, SVHN, Tiny ImageNet, LSUN, Places365) CIFAR100-LT OOD tests (Textures, SVHN, Tiny ImageNet, LSUN, Places365) ImageNet-LT with ImageNet-1k-OOD
Metrics: AUROC AUPR FPR95 ACC95 ACC@FPRn ACC@TPRn
Baselines: OE PASCL EnergyOE OECC ST (MSP) OLTR RIDE GLMC
Builds on: Outlier Exposure PASCL ATOM CutMix Logit adjustment loss
Closest papers by meaning
- Long-Tailed Out-of-Distribution Detection: Prioritizing Attention to Tail (AAAI 2025 · similarity 0.95)
- Out-of-Distribution Detection in Long-Tailed Recognition with Calibrated Outlier Class Learning (AAAI 2024 · similarity 0.93)
- Long-Tailed Out-of-Distribution Detection via Normalized Outlier Distribution Adaptation (NeurIPS 2024 · similarity 0.93)
- Is Complex Training Necessary for Long-Tailed OOD Detection? A Re-think from Feature Geometry (NeurIPS 2026 · similarity 0.92)
- Rethinking Out-of-Distribution Detection on Imbalanced Data Distribution (NeurIPS 2024 · similarity 0.91)
- Image Background Serves as Good Proxy for Out-of-distribution Data (ICLR 2024 · similarity 0.90)
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.