Research map / Label-efficient and robust learning / OOD and anomaly detection
Long-Tailed Out-of-Distribution Detection via Normalized Outlier Distribution Adaptation
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 AdaptOD, which adapts a learned outlier energy distribution to the true OOD distribution at test time (DODA) and trains with a dual-normalized energy loss (DNE) for long-tailed OOD detection.
Problem
OOD detection under long-tailed ID training data, where auxiliary outliers have a distribution shift from true OOD samples, especially for head vs. tail ID classes.
Approach
Combines a dual-normalized energy loss (class-wise and sample-wise batch-normalized energy, margin-free) for learning a better vanilla outlier distribution with dynamic test-time adaptation that momentum-updates the outlier distribution using samples predicted as OOD, calibrating the global energy score without retraining or memory banks.
Key results
- On CIFAR10-LT, AdaptOD reaches 94.69 AUC and 27.26 FPR95 versus 93.28 AUC and 30.88 FPR95 for COCL (averaged over six OOD sets).
- On CIFAR100-LT, AdaptOD reaches 81.93 AUC versus 78.25 for COCL; on ImageNet-LT it reaches 74.32 AUC versus 71.85 for COCL.
- Plugging DODA into OE, EnergyOE, BERL and COCL consistently improves all four OOD metrics, and AdaptOD comes close to an oracle that uses true OOD labels for adaptation (CIFAR10-LT AUC 94.69 vs 95.33).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add DODA to the EnergyOE baseline | EnergyOE · CIFAR10-LT, six-OOD average | AUC: 91.92 -> 93.36 | Helped |
| add DODA to the EnergyOE baseline | EnergyOE · CIFAR100-LT, six-OOD average | AUC: 76.40 -> 79.40 | Helped |
| add DODA to the EnergyOE baseline | EnergyOE · ImageNet-LT, ImageNet-1k-OOD | AUC: 69.43 -> 73.12 | Helped |
| add DNE-C and DNE-S to EnergyOE training (ablation) | AdaptOD without DODA · CIFAR10-LT, six-OOD average | AUC: 91.92 -> 92.77 | Helped |
| add DNE-C and DNE-S to EnergyOE training (ablation) | AdaptOD without DODA · CIFAR100-LT, six-OOD average | AUC: 76.40 -> 77.92 | Helped |
| add DNE-C and DNE-S to EnergyOE training (ablation) | AdaptOD without DODA · ImageNet-LT, ImageNet-1k-OOD | AUC: 69.43 -> 72.04 | Helped |
| add DODA, DNE-C, and DNE-S together (ablation) | EnergyOE baseline · CIFAR10-LT, six-OOD average | AUC: 91.92 -> 94.69 | Helped |
| add DODA, DNE-C, and DNE-S together (ablation) | EnergyOE baseline · CIFAR100-LT, six-OOD average | AUC: 76.40 -> 81.93 | Helped |
| add DODA, DNE-C, and DNE-S together (ablation) | EnergyOE baseline · ImageNet-LT, ImageNet-1k-OOD | AUC: 69.43 -> 74.32 | Helped |
| use AdaptOD instead of COCL | long-tailed OOD detection · CIFAR10-LT, six-OOD average | AUC: 93.28 -> 94.69 | Helped |
| use AdaptOD instead of COCL | long-tailed OOD detection · CIFAR100-LT, six-OOD average | AUC: 78.25 -> 81.93 | Helped |
| use AdaptOD instead of COCL | long-tailed OOD detection · ImageNet-LT, ImageNet-1k-OOD | AUC: 71.85 -> 74.32 | Helped |
Surprising findings
- Test-time adaptation with predicted OOD samples nearly matches an oracle that uses ground-truth OOD labels for updating the outlier distribution.
Limitations
- Relies on the OOD filter threshold (alpha=3 default) and the accuracy of predicted OOD samples during inference.
- Gains over strong baselines are modest on some datasets, e.g. CIFAR10-LT average AUC improves by about 1.4 points over COCL.
Experimental setup
Datasets: CIFAR10-LT CIFAR100-LT ImageNet-LT TinyImages80M ImageNet-Extra CIFAR Texture SVHN LSUN Places365 TinyImageNet ImageNet-1k-OOD
Benchmarks: CIFAR10-LT OOD (SC-OOD six datasets) CIFAR100-LT OOD (SC-OOD six datasets) ImageNet-LT with ImageNet-1k-OOD
Metrics: AUC AP-in AP-out FPR95 ACC
Baselines: OE EnergyOE PASCL EAT Class Prior BERL COCL AUTO AdaODD
Builds on: energy-based OOD detection outlier exposure EnergyOE COCL AUTO
Closest papers by meaning
- Long-Tailed Out-of-Distribution Detection: Prioritizing Attention to Tail (AAAI 2025 · similarity 0.93)
- Out-of-Distribution Detection in Long-Tailed Recognition with Calibrated Outlier Class Learning (AAAI 2024 · similarity 0.93)
- EAT: Towards Long-Tailed Out-of-Distribution Detection (AAAI 2024 · similarity 0.93)
- Rethinking Out-of-Distribution Detection on Imbalanced Data Distribution (NeurIPS 2024 · similarity 0.91)
- Expecting The Unexpected: Towards Broad Out-Of-Distribution Detection (NeurIPS 2024 · similarity 0.90)
- Out-Of-Distribution Detection with Diversification (Provably) (NeurIPS 2024 · 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.