Out-of-Distribution Detection on CIFAR100 (ID) vs SVHN (OOD) (test)
98.7AUROCKernel PCA (CoRP)
Evaluation results
| Method | Links | |||
|---|---|---|---|---|
| Kernel PCA (CoRP)Architecture=ResNetV2-50 (BiT-M), Labels=false2025.05 | 98.7 | 6.17 | — | |
| DICEArchitecture=ResNetV2-50 (BiT-M), Labels=false2025.05 | 97.5 | 13.92 | — | |
| GAIA-ZArchitecture=ResNet-34 (TIMM), Labels=false2025.05 | 97.34 | 13.33 | — | |
| KNNArchitecture=ResNetV2-50 (BiT-M), Labels=false2025.05 | 96.8 | 14.01 | — | |
| GradPCAArchitecture=ResNetV2-50 (BiT-M), Labels=true2025.05 | 96.58 | 17.2 | — | |
| GradPCA+DICEArchitecture=ResNetV2-50 (BiT-M), Labels=true2025.05 | 96.57 | 18.11 | — | |
| DMD-aBackbone=ViTB162025.12 | 95 | — | — | |
| Revisited PCAArchitecture=ResNetV2-50 (BiT-M), Labels=false2025.05 | 94.85 | 25.62 | — | |
| FSBackbone=ViTB162025.12 | 94 | — | — | |
| MahalanobisArchitecture=ResNetV2-50 (BiT-M), Labels=true2025.05 | 93.74 | 38.07 | — | |
| Proj. GradsArchitecture=ResNetV2-50 (BiT-M), Labels=true2025.05 | 93.01 | 36.78 | — | |
| GradOrthArchitecture=ResNetV2-50 (BiT-M), Labels=false2025.05 | 92.97 | 26.974 | — | |
| DOCBackbone=ViTB162025.12 | 91 | — | — | |
| DPSBackbone=ViT-B/162025.12 | 90.71 | — | — | |
| MACSBackbone=ViTB162025.12 | 90 | — | — | |
| MSPBackbone=ViTB162025.12 | 90 | — | — | |
| DMD-bBackbone=ViTB162025.12 | 90 | — | — | |
| LoRA-EnsembleEnsemble size=16, LoRA=true2024.05 | 89.9 | 41.6 | 95.2 | |
| NUQBackbone=ResNet-50, Normalization=Spectral2022.02 | 89.7 | — | — | |
| DDUBackbone=ResNet-50, Normalization=Spectral2022.02 | 89.6 | — | — | |
| DLBBackbone=ViT-B/162025.12 | 89.52 | — | — | |
| Vanilla HNNScore function=Energy score2021.07 | 89.43 | 55.44 | 97.69 | |
| TEBackbone=ViT-B/162025.12 | 89.12 | — | — | |
| GAIA-AArchitecture=ResNetV2-50 (BiT-M), Labels=false2025.05 | 89.12 | 52.74 | — | |
| GradPCAArchitecture=ResNet-34 (TIMM), Labels=true2025.05 | 89.1 | 61.22 | — | |
| ReActArchitecture=ResNetV2-50 (BiT-M), Labels=false2025.05 | 89.03 | 44.77 | — | |
| PS-KDBackbone=ViT-B/162025.12 | 88.82 | — | — | |
| DUQ end-to-endBackbone=ResNet-50, Training=End-to-end2022.02 | 88.7 | — | — | |
| BaselineBackbone=ViT-B/162025.12 | 88.42 | — | — | |
| DMD-aBackbone=VGG162025.12 | 88 | — | — | |
| NCI (w/o filter)Architecture=ResNetV2-50 (BiT-M), Labels=false2025.05 | 87.54 | 76.36 | — | |
| EnergyArchitecture=ResNetV2-50 (BiT-M), Labels=false2025.05 | 87.44 | 63.74 | — | |
| Max logitsArchitecture=ResNetV2-50 (BiT-M), Labels=false2025.05 | 87.17 | 65.17 | — | |
| ODINArchitecture=ResNetV2-50 (BiT-M), Labels=false2025.05 | 87.17 | 65.18 | — | |
| ENNScore function=Energy score2021.07 | 86.37 | 53.37 | 96.78 | |
| FSBackbone=VGG162025.12 | 86 | — | — | |
| Single Network with LoRAEnsemble size=1, LoRA=true2024.05 | 85.9 | 49.7 | 93.1 | |
| MSPArchitecture=ResNetV2-50 (BiT-M), Labels=false2025.05 | 85.87 | 68.42 | — | |
| HMCOOD score=Maximum Softmax Probability (MSP)2024.08 | 85.8 | — | — | |
| Clipped HNNsScore function=Energy score2021.07 | 84.41 | 84.12 | 96.72 | |
| Max logitsArchitecture=ResNet-34 (TIMM), Labels=false2025.05 | 83.83 | 71.38 | — | |
| DUQ HeadBackbone=ResNet-50, Training=Pre-trained feature extractor2022.02 | 83.6 | — | — | |
| EnsemblesBackbone=ResNet-502022.02 | 82.9 | — | — | |
| DLBBackbone=ResNet-502025.12 | 82.51 | — | — | |
| DPSBackbone=DenseNet-1692025.12 | 82.41 | — | — | |
| MSPBackbone=VGG162025.12 | 82 | — | — | |
| DOCBackbone=VGG162025.12 | 82 | — | — | |
| DMD-uBackbone=VGG162025.12 | 82 | — | — | |
| GAIA-ZArchitecture=ResNetV2-50 (BiT-M), Labels=false2025.05 | 81.78 | 79.76 | — | |
| TTABackbone=ResNet-50, Protocol=Test-Time Augmentation2022.02 | 81.6 | — | — | |
| Split-EnsembleEnsemble size=162024.05 | 81.2 | 75 | 69.9 | |
| MACSBackbone=VGG162025.12 | 81 | — | — | |
| DLBBackbone=DenseNet-1692025.12 | 80.55 | — | — | |
| ENN (EpiNet)Ensemble size=162024.05 | 78.6 | 50.5 | 88 | |
| L2EOOD score=Maximum Softmax Probability (MSP)2024.08 | 78.6 | — | — | |
| DPSBackbone=ResNet-502025.12 | 78.19 | — | — | |
| BaselineBackbone=DenseNet-1692025.12 | 77.92 | — | — | |
| PS-KDBackbone=ResNet-502025.12 | 76.66 | — | — | |
| Single NetworkEnsemble size=12024.05 | 76.4 | 55.9 | 86.8 | |
| Explicit EnsembleEnsemble size=162024.05 | 74.8 | 61 | 86.6 | |
| TEBackbone=ResNet-502025.12 | 74.68 | — | — | |
| BaselineBackbone=ResNet-502025.12 | 74.35 | — | — | |
| PS-KDBackbone=DenseNet-1692025.12 | 73.95 | — | — | |
| TEBackbone=DenseNet-1692025.12 | 68.97 | — | — | |
| Rel-UBackbone=ViTB162025.12 | 67 | — | — | |
| Rel-UBackbone=VGG162025.12 | 64 | — | — | |
| CSGMCMCOOD score=Maximum Softmax Probability (MSP)2024.08 | 55.5 | — | — | |
| DEOOD score=Maximum Softmax Probability (MSP)2024.08 | 54.5 | — | — | |
| MC DropoutEnsemble size=162024.05 | 52.3 | 94.8 | 74.5 | |
| DMD-uBackbone=ViTB162025.12 | 51 | — | — | |
| DMD-bBackbone=VGG162025.12 | 45 | — | — |