Research map / Label-efficient and robust learning / OOD and anomaly detection
Out-of-Distribution Detection in Long-Tailed Recognition with Calibrated Outlier Class Learning
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 COCL, which learns an explicit outlier class for OOD detection in long-tailed recognition, adding debiased large margin losses and outlier-class-aware logit calibration to reduce head/tail-OOD confusion.
Problem
OOD detection under long-tailed recognition, where OOD samples are misclassified as head classes and tail-class samples are mistaken for OOD samples; prior outlier-exposure methods need an accurate prior over long-tailed ID data, which is hard to obtain.
Approach
Replace the uniform/prior-based outlier exposure target with a learned (k+1)-th outlier class trained on auxiliary OOD data, add a tail-prototype contrastive loss and a head-class large-margin loss to debias the outlier class, and calibrate logits with prior counts including the outlier class (n_{k+1}=1) at inference.
Key results
- On CIFAR10-LT (ResNet18), COCL reaches average AUC 93.28, FPR 30.88 and ACC 81.56 across six OOD test sets, versus 89.76 AUC for OE and 91.56 AUC for OCL+LA.
- On CIFAR100-LT, COCL achieves AUC 78.25 and ACC 46.41, outperforming BERL (77.75 AUC, 45.88 ACC).
- On ImageNet-LT, COCL achieves AUC 71.85, FPR 75.60 and ACC 51.11, the best among compared methods including BERL (71.16 AUC).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| replace OE with outlier-class learning | LTR OOD detection · CIFAR10-LT, six OOD datasets average | FPR: 53.19 -> 41.13 | Helped |
| add calibrated outlier-class learning (COCL) over OCL | LTR OOD detection · CIFAR10-LT, six OOD datasets average | AUC: 89.91 -> 93.28 | Helped |
| add calibrated outlier-class learning (COCL) over OCL | LTR OOD detection · CIFAR100-LT, six OOD datasets average | AUC: 73.56 -> 78.25 | Helped |
| add calibrated outlier-class learning (COCL) over OCL | LTR OOD detection · ImageNet-LT, ImageNet-1k-OOD | AUC: 68.67 -> 71.85 | Helped |
| add tail-class prototype learning (TCPL) (ablation) | COCL · CIFAR10-LT | AUC: 89.91 -> 91.23 | Helped |
| add debiased head-class learning (DHCL) (ablation) | COCL · CIFAR10-LT | AUC: 89.91 -> 91.08 | Helped |
| add outlier-class-aware logit calibration (OLC) (ablation) | COCL · CIFAR10-LT ID | ACC: 74.48 -> 79.40 | Helped |
| add tail-class prototype learning (TCPL) (ablation) | COCL · CIFAR100-LT | AUC: 73.56 -> 75.14 | Helped |
| add debiased head-class learning (DHCL) (ablation) | COCL · CIFAR100-LT | AUC: 73.56 -> 74.70 | Helped |
| add outlier-class-aware logit calibration (OLC) (ablation) | COCL · CIFAR100-LT ID | ACC: 41.54 -> 45.62 | Helped |
| add tail-class prototype learning (TCPL) (ablation) | COCL · ImageNet-LT | AUC: 68.67 -> 70.08 | Helped |
| add debiased head-class learning (DHCL) (ablation) | COCL · ImageNet-LT | AUC: 68.67 -> 69.64 | Helped |
Surprising findings
- OCL works much less effectively than OE on CIFAR100-LT when the OOD test set is CIFAR10, where OE slightly beats COCL.
Limitations
- OCL underperforms OE on CIFAR100-LT when the OOD set is CIFAR10 due to semantic similarity between the datasets.
- Requires auxiliary OOD data (TinyImages, ImageNet-22k) for outlier class training.
- Several hyperparameters (gamma, alpha, beta, margin) are sensitive and must be tuned.
Experimental setup
Datasets: CIFAR10-LT CIFAR100-LT ImageNet-LT TinyImages 80M ImageNet-Extra
Benchmarks: SC-OOD benchmark (CIFAR10, Texture, SVHN, LSUN, Places365, TinyImageNet) ImageNet-1k-OOD CIFAR10-LT CIFAR100-LT ImageNet-LT
Metrics: AUC AP-in AP-out FPR95 ACC
Baselines: MSP OE EnergyOE PASCL Open Sampling Class Prior BERL OCL
Builds on: outlier exposure PASCL logit adjustment SC-OOD benchmark
Closest papers by meaning
- EAT: Towards Long-Tailed Out-of-Distribution Detection (AAAI 2024 · similarity 0.93)
- Long-Tailed Out-of-Distribution Detection via Normalized Outlier Distribution Adaptation (NeurIPS 2024 · similarity 0.93)
- Long-Tailed Out-of-Distribution Detection: Prioritizing Attention to Tail (AAAI 2025 · similarity 0.92)
- Rethinking Out-of-Distribution Detection on Imbalanced Data Distribution (NeurIPS 2024 · similarity 0.89)
- A Noisy Elephant in the Room: Is Your Out-of-Distribution Detector Robust to Label Noise? (CVPR 2024 · similarity 0.89)
- Is Complex Training Necessary for Long-Tailed OOD Detection? A Re-think from Feature Geometry (NeurIPS 2026 · similarity 0.89)
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.