Research map / Label-efficient and robust learning / OOD and anomaly detection
Enhancing the Power of OOD Detection via Sample-Aware Model Selection
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 ZODE, which applies Benjamini-Hochberg p-value model selection over a zoo of pre-trained models so each test input is flagged OOD if any selected model detects it, improving post hoc OOD detection on CIFAR10 and ImageNet.
Problem
Post hoc OOD detection performance depends heavily on which pre-trained model is chosen, and no guidance exists for selecting one in advance; a naive ensemble over models fails to keep the ID true positive rate.
Approach
Normalize each model's detection score into a p-value and apply the Benjamini-Hochberg procedure per test sample to select which zoo models flag the input as OOD, classifying as ID when none do, which keeps the TPR at the target level while exploiting model complementarity.
Key results
- ZODE-KNN reduces average FPR on CIFAR10 benchmarks from 11.07% (KNN+) to 3.83%, a 65.40% relative improvement.
- On ImageNet-1K, ZODE-KNN reduces average FPR from 38.47% (KNN+) to 24.14%, a 37.25% relative improvement.
- On CIFAR10 vs CIFAR100, ZODE-KNN achieves FPR 18.29% vs 38.50% for SSD+.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| apply sample-aware model selection across the model zoo instead of using KNN+ alone | KNN OOD detection · CIFAR10 OOD benchmarks (average) | FPR: 11.07 -> 3.83 | Helped |
| apply sample-aware model selection across the model zoo instead of using SSD+ | CIFAR10 vs CIFAR100 OOD detection · CIFAR10 vs CIFAR100 | FPR: 38.50 -> 18.29 | Helped |
| apply sample-aware model selection across the model zoo instead of using KNN+ alone | KNN OOD detection · ImageNet-1K OOD benchmarks (average) | FPR: 38.47 -> 24.14 | Helped |
| replace ZODE with the naive any-model ensemble (ablation) | OOD detection on ImageNet-1K · ImageNet-1K OOD benchmarks (average) | FPR: 10.42 vs 24.14 | Helped |
| replace ZODE with average p-value aggregation (ablation) | OOD detection on ImageNet-1K · ImageNet-1K OOD benchmarks (average) | FPR: 44.42 vs 24.14 | Hurt |
| replace ZODE with 60% voting across models (ablation) | OOD detection on ImageNet-1K · ImageNet-1K OOD benchmarks (average) | FPR: 29.84 vs 24.14 | Hurt |
| replace ZODE with Benjamini-Yekutieli multiple-testing correction (ablation) | OOD detection on ImageNet-1K · ImageNet-1K OOD benchmarks (average) | FPR: 42.21 vs 24.14 | Hurt |
| use a model zoo with sample-aware selection rather than the best single-model KNN detector | KNN OOD detection · CIFAR10 OOD benchmarks (average) | FPR: 11.03 -> 3.83 | Helped |
| use a model zoo with sample-aware selection rather than the best single-model KNN detector | KNN OOD detection · ImageNet-1K OOD benchmarks (average) | FPR: 26.97 -> 24.14 | Helped |
Surprising findings
- The naive ensemble that flags an input as OOD if any model does fails to maintain the target TPR, with error accumulating as the number of models grows.
Limitations
- Test-time storage cost is higher because p-values require stored validation-sample scores.
- Theoretical TPR guarantee assumes independence across models, which often fails in practice; empirical TPR is validated instead.
Experimental setup
Datasets: CIFAR10 CIFAR100 ImageNet-1K
Benchmarks: CIFAR10 vs SVHN CIFAR10 vs LSUN CIFAR10 vs iSUN CIFAR10 vs Texture CIFAR10 vs Places365 CIFAR10 vs CIFAR100 ImageNet-1K vs iNaturalist ImageNet-1K vs SUN ImageNet-1K vs Places ImageNet-1K vs Textures
Metrics: FPR at 95% TPR AUROC TPR
Baselines: MSP ODIN Energy GODIN Mahalanobis KNN SSD+ KNN+ GRAM MaSF CSI Naive ensemble
Builds on: KNN OOD detection Benjamini-Hochberg procedure Mahalanobis detector Energy score MSP baseline
Closest papers by meaning
- Beyond Clean Training Data: A Versatile and Model-Agnostic Framework for Out-of-Distribution Detection with Contaminated Training Data (CVPR 2025 · similarity 0.91)
- The Best of Both Worlds: On the Dilemma of Out-of-distribution Detection (NeurIPS 2024 · similarity 0.90)
- Fast Decision Boundary based Out-of-Distribution Detector (ICML 2024 · similarity 0.90)
- Leveraging Perturbation Robustness to Enhance Out-of-Distribution Detection (CVPR 2025 · similarity 0.89)
- What If the Input is Expanded in OOD Detection? (NeurIPS 2024 · similarity 0.89)
- Image Background Serves as Good Proxy for Out-of-distribution Data (ICLR 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.