OOD Detection on CIFAR-10 (ID) vs Places 365 (OOD)
99.6AUROCMahalanobis
Evaluation results
| Method | Links | ||
|---|---|---|---|
| MahalanobisBackbone=ResNetV2-50 (BiT-M)2025.05 | 99.6 | 1.64 | |
| GradPCA-BatchBackbone=ResNetV2-50 (BiT-M)2025.05 | 99.3 | 3.2 | |
| MedixTraining Data=Pin and Pwild2025.10 | 99.1 | 2.98 | |
| KNNBackbone=ResNetV2-50 (BiT-M)2025.05 | 98.74 | 5.52 | |
| Hopfield BoostingBackbone=ResNet-182024.05 | 98.51 | 4.28 | |
| Hopfield BoostingMethod type=OE, Augmentations=Weak, Auxiliary outlier data=true, Backbone=ResNet-182024.05 | 98.51 | 4.28 | |
| WOODSTraining Data=Pin and Pwild2025.10 | 98.05 | 10.19 | |
| POEMBackbone=ResNet-182024.05 | 97.56 | 7.7 | |
| GradPCA-VecBackbone=ResNetV2-50 (BiT-M)2025.05 | 97.47 | 11.01 | |
| Kernel PCA (CoRP)Backbone=ResNetV2-50 (BiT-M)2025.05 | 97.31 | 13.29 | |
| GradPCA (block 4)Backbone=ResNetV2-50 (BiT-M)2025.05 | 97.25 | 10.96 | |
| DivOEBackbone=ResNet-182024.05 | 96.95 | 13.7 | |
| MixOEBackbone=ResNet-182024.05 | 96.92 | 16.3 | |
| KNNContrastive Learning=No2026.01 | 96.84 | 18.5 | |
| DALBackbone=ResNet-182024.05 | 96.77 | 14.22 | |
| DOSBackbone=ResNet-182024.05 | 96.63 | 12.26 | |
| GAIA-ABackbone=ResNetV2-50 (BiT-M)2025.05 | 96.51 | 18.16 | |
| EBO-OEBackbone=ResNet-182024.05 | 96.39 | 11.77 | |
| Energy (w/ OE)Training Data=Pin and Pwild2025.10 | 96.18 | 14.66 | |
| MSP-OEBackbone=ResNet-182024.05 | 95.91 | 21.42 | |
| KNNTraining Data=Pin only2025.10 | 95.69 | 25.29 | |
| SSD+Contrastive Learning=Yes2026.01 | 95.57 | 22.05 | |
| GradPCA (block 3)Backbone=ResNet-34 (TIMM)2025.05 | 95.33 | 24.87 | |
| DOEBackbone=ResNet-182024.05 | 95.06 | 19.72 | |
| PALMMethod type=Training, Augmentations=Strong, Auxiliary outlier data=false, Backbone=ResNet-182024.05 | 94.95 | 22.97 | |
| SSD+Method type=Training, Augmentations=Strong, Auxiliary outlier data=false, Backbone=ResNet-182024.05 | 94.93 | 48.03 | |
| KNN+Contrastive Learning=Yes2026.01 | 94.88 | 23.05 | |
| OETraining Data=Pin and Pwild2025.10 | 94.88 | 19.48 | |
| KNN+Training Data=Pin only2025.10 | 94.84 | 24.69 | |
| GAIA-ABackbone=ResNet-34 (TIMM)2025.05 | 94.44 | 29.54 | |
| GradPCABackbone=ResNetV2-50 (BiT-M)2025.05 | 94.41 | 23.16 | |
| NCIBackbone=ResNetV2-50 (BiT-M)2025.05 | 94.35 | 16.14 | |
| NGCSupervised detection=false, IND noise level=50% sym., OOD samples in training set=20k, OOD samples in test set=10k2021.08 | 94.31 | — | |
| NCIBackbone=ResNet-34 (TIMM)2025.05 | 94.26 | 24.56 | |
| CIDERContrastive Learning=Yes2026.01 | 94.09 | 23.88 | |
| NPOSMethod type=Training, Augmentations=Strong, Auxiliary outlier data=false, Backbone=ResNet-182024.05 | 93.76 | 24.99 | |
| CSITraining Data=Pin only2025.10 | 93.64 | 34.95 | |
| GradPCA+DICEBackbone=ResNetV2-50 (BiT-M)2025.05 | 93.35 | 28.77 | |
| CSIContrastive Learning=Yes2026.01 | 93.04 | 38.31 | |
| CreDROEnsemble Number M=202026.02 | 92.7 | — | |
| CreDRO2026.02 | 92.7 | — | |
| Kernel PCA (CoRP)Backbone=ResNet-34 (TIMM)2025.05 | 92.51 | 41.7 | |
| CreDROEnsemble Number M=152026.02 | 92.5 | — | |
| GradPCABackbone=ResNet-34 (TIMM)2025.05 | 92.29 | 34.75 | |
| KNNBackbone=ResNet-34 (TIMM)2025.05 | 92.22 | 42.12 | |
| CreDROEnsemble Number M=102026.02 | 92.2 | — | |
| ProxyAnchorContrastive Learning=Yes2026.01 | 92.06 | 43.46 | |
| CreDEEnsemble Number M=152026.02 | 91.8 | — | |
| CreRL1.0Ensemble Number M=202026.02 | 91.8 | — | |
| CreDEEnsemble Number M=202026.02 | 91.8 | — | |
| CreRLparameter=1.02026.02 | 91.8 | — | |
| CreDE2026.02 | 91.8 | — | |
| CreRL1.0Ensemble Number M=152026.02 | 91.7 | — | |
| GAIA-ZBackbone=ResNet-34 (TIMM)2025.05 | 91.7 | 32.34 | |
| CreDROEnsemble Number M=52026.02 | 91.6 | — | |
| CreDEEnsemble Number M=102026.02 | 91.6 | — | |
| CreWraEnsemble Number M=202026.02 | 91.6 | — | |
| CreWra2026.02 | 91.6 | — | |
| CreWraEnsemble Number M=152026.02 | 91.5 | — | |
| CreRL1.0Ensemble Number M=102026.02 | 91.3 | — | |
| CreEns0.0Ensemble Number M=202026.02 | 91.3 | — | |
| CreEnsparameter=0.02026.02 | 91.3 | — | |
| Comp.VAEContrastive Learning=No2026.01 | 91.3 | 20.4 | |
| GradPCA-VecBackbone=ResNet-34 (TIMM)2025.05 | 91.17 | 49.1 | |
| GradPCA-BatchBackbone=ResNet-34 (TIMM)2025.05 | 91.14 | 48.65 | |
| GradOrthBackbone=ResNet-34 (TIMM)2025.05 | 91.13 | 37.6 | |
| CreWraEnsemble Number M=102026.02 | 91.1 | — | |
| CreEns0.0Ensemble Number M=152026.02 | 91.1 | — | |
| EN-DROEnsemble Number M=202026.02 | 91.1 | — | |
| EN-DRO2026.02 | 91.1 | — | |
| EnergyContrastive Learning=No2026.01 | 91.02 | 42.77 | |
| EN-DROEnsemble Number M=152026.02 | 91 | — | |
| ODINContrastive Learning=No2026.01 | 90.98 | 43.4 | |
| Revisited PCABackbone=ResNet-34 (TIMM)2025.05 | 90.83 | 35.81 | |
| EN-DROEnsemble Number M=102026.02 | 90.8 | — | |
| CreEns0.0Ensemble Number M=102026.02 | 90.8 | — | |
| CreDEEnsemble Number M=52026.02 | 90.6 | — | |
| ReActTraining Data=Pin only2025.10 | 90.44 | 41.44 | |
| CreWraEnsemble Number M=52026.02 | 90.4 | — | |
| CreRL1.0Ensemble Number M=52026.02 | 90.4 | — | |
| EN-DROEnsemble Number M=52026.02 | 90.3 | — | |
| GradPCA+DICEBackbone=ResNet-34 (TIMM)2025.05 | 90.26 | 39.54 | |
| DEEnsemble Number M=202026.02 | 90 | — | |
| DE2026.02 | 90 | — | |
| CreEns0.0Ensemble Number M=52026.02 | 89.9 | — | |
| DEEnsemble Number M=152026.02 | 89.9 | — | |
| EnergyTraining Data=Pin only2025.10 | 89.89 | 40.14 | |
| DEEnsemble Number M=102026.02 | 89.6 | — | |
| MahalanobisBackbone=ResNet-34 (TIMM)2025.05 | 89.2 | 56.02 | |
| DEEnsemble Number M=52026.02 | 89.1 | — | |
| GENMethod type=Post-hoc, Augmentations=Weak, Auxiliary outlier data=false, Backbone=ResNet-182024.05 | 88.68 | 44.87 | |
| MSPContrastive Learning=No2026.01 | 88.64 | 62.46 | |
| EBOMethod type=Post-hoc, Augmentations=Weak, Auxiliary outlier data=false, Backbone=ResNet-182024.05 | 88.53 | 45.63 | |
| CreBNN2026.02 | 88.5 | — | |
| MaxLogitMethod type=Post-hoc, Augmentations=Weak, Auxiliary outlier data=false, Backbone=ResNet-182024.05 | 88.42 | 45.63 | |
| Proj. GradsBackbone=ResNetV2-50 (BiT-M)2025.05 | 88.36 | 36.52 | |
| ASHTraining Data=Pin only2025.10 | 88.34 | 48.45 | |
| MSPTraining Data=Pin only2025.10 | 88.2 | 59.48 | |
| MSPMethod type=Post-hoc, Augmentations=Weak, Auxiliary outlier data=false, Backbone=ResNet-182024.05 | 88.06 | 49.21 | |
| MSPBackbone=ResNet-34 (TIMM)2025.05 | 87.5 | 54.45 |