OOD Detection on Textures (OOD) vs CIFAR-10 (ID) (test)
0.02FPR@95Mahalanobis
Evaluation results
| Method | Links | ||
|---|---|---|---|
| MahalanobisBackbone=ResNetV2-50 (BiT-M)2025.05 | 0.02 | 99.97 | |
| GradPCA-BatchBackbone=ResNetV2-50 (BiT-M)2025.05 | 0.02 | 99.99 | |
| Kernel PCA (CoRP)Backbone=ResNetV2-50 (BiT-M)2025.05 | 0.12 | 99.96 | |
| KNNBackbone=ResNetV2-50 (BiT-M)2025.05 | 0.18 | 99.94 | |
| GradPCA-VecBackbone=ResNetV2-50 (BiT-M)2025.05 | 0.36 | 99.91 | |
| GradPCA (block 4)Backbone=ResNetV2-50 (BiT-M)2025.05 | 0.62 | 99.87 | |
| GradPCA+DICEBackbone=ResNetV2-50 (BiT-M)2025.05 | 1.17 | 99.67 | |
| GradPCABackbone=ResNetV2-50 (BiT-M)2025.05 | 1.33 | 99.73 | |
| NCIBackbone=ResNetV2-50 (BiT-M)2025.05 | 2.27 | 99.04 | |
| GAIA-ZBackbone=ResNet-34 (TIMM)2025.05 | 9.02 | 98.21 | |
| GAIA-ZBackbone=ResNetV2-50 (BiT-M)2025.05 | 9.13 | 98 | |
| GAIA-ABackbone=ResNetV2-50 (BiT-M)2025.05 | 10.74 | 97.87 | |
| Proj. GradsBackbone=ResNetV2-50 (BiT-M)2025.05 | 12.52 | 96.66 | |
| Revisited PCABackbone=ResNetV2-50 (BiT-M)2025.05 | 15.91 | 94.58 | |
| Hopfield BoostingBackbone=ResNet-182024.05 | 16 | 99.84 | |
| Hopfield BoostingMethod type=OE, Augmentations=Weak, Auxiliary outlier data=true, Backbone=ResNet-182024.05 | 16 | 99.84 | |
| GAIA-ABackbone=ResNet-34 (TIMM)2025.05 | 16.76 | 96.79 | |
| MSPBackbone=ResNetV2-50 (BiT-M)2025.05 | 18.16 | 93.12 | |
| FedRoD + FOOGDalpha=0.12024.10 | 19.46 | 93.39 | |
| NCIBackbone=ResNet-34 (TIMM)2025.05 | 20.46 | 94.64 | |
| GradOrthBackbone=DenseNet, In-Distribution Dataset=CIFAR-102023.08 | 20.63 | 94.77 | |
| MahalanobisBackbone=DenseNet, In-Distribution Dataset=CIFAR-102023.08 | 21.51 | 92.15 | |
| ODINBackbone=ResNetV2-50 (BiT-M)2025.05 | 24.33 | 89.66 | |
| ASH-SBackbone=DenseNet, In-Distribution Dataset=CIFAR-102023.08 | 24.34 | 95.09 | |
| Max logitsBackbone=ResNetV2-50 (BiT-M)2025.05 | 24.34 | 89.65 | |
| GradPCA (block 3)Backbone=ResNet-34 (TIMM)2025.05 | 25.78 | 95.86 | |
| FedAvg + FOOGDalpha=0.12024.10 | 28.6 | 91.75 | |
| EnergyBackbone=ResNetV2-50 (BiT-M)2025.05 | 30.42 | 87.53 | |
| VRA-PBackbone=DenseNet, In-Distribution Dataset=CIFAR-102023.08 | 34.89 | 93.42 | |
| ASH-BBackbone=DenseNet, In-Distribution Dataset=CIFAR-102023.08 | 35.73 | 92.88 | |
| DICEBackbone=DenseNet, In-Distribution Dataset=CIFAR-102023.08 | 41.9 | 93.36 | |
| KNNBackbone=ResNet-34 (TIMM)2025.05 | 42.03 | 92.77 | |
| GradPCA-VecBackbone=ResNet-34 (TIMM)2025.05 | 42.92 | 92.64 | |
| GradPCA-BatchBackbone=ResNet-34 (TIMM)2025.05 | 43.29 | 92.46 | |
| Kernel PCA (CoRP)Backbone=ResNet-34 (TIMM)2025.05 | 43.43 | 93.15 | |
| ReActBackbone=DenseNet, In-Distribution Dataset=CIFAR-102023.08 | 43.58 | 92.47 | |
| Revisited PCABackbone=ResNet-34 (TIMM)2025.05 | 45.35 | 87.51 | |
| EnergyBackbone=ResNet-34 (TIMM)2025.05 | 47.75 | 83.95 | |
| Max logitsBackbone=ResNet-34 (TIMM)2025.05 | 48.6 | 83.94 | |
| ODINBackbone=ResNet-34 (TIMM)2025.05 | 48.6 | 83.94 | |
| POEMBackbone=ResNet-182024.05 | 49 | 99.72 | |
| GradPCABackbone=ResNet-34 (TIMM)2025.05 | 49.22 | 91.54 | |
| GradNormBackbone=DenseNet, In-Distribution Dataset=CIFAR-102023.08 | 50.26 | 89.72 | |
| ReActBackbone=ResNet-34 (TIMM)2025.05 | 50.37 | 82.23 | |
| GradOrthBackbone=ResNetV2-50 (BiT-M)2025.05 | 50.5 | 90.83 | |
| ASH-PBackbone=DenseNet, In-Distribution Dataset=CIFAR-102023.08 | 50.85 | 88.29 | |
| FedICONalpha=0.12024.10 | 51.57 | 80.96 | |
| MahalanobisBackbone=ResNet-34 (TIMM)2025.05 | 51.6 | 91.31 | |
| DICEBackbone=ResNet-34 (TIMM)2025.05 | 52.31 | 79.5 | |
| FedRoDalpha=0.12024.10 | 53.24 | 81.52 | |
| GradOrthBackbone=ResNet-34 (TIMM)2025.05 | 53.5 | 89.8 | |
| FedTHEalpha=0.12024.10 | 53.58 | 82.19 | |
| FOSTERalpha=0.12024.10 | 54.23 | 77.62 | |
| GradPCA+DICEBackbone=ResNet-34 (TIMM)2025.05 | 55.38 | 88.21 | |
| Energy scoreBackbone=DenseNet, In-Distribution Dataset=CIFAR-102023.08 | 56.12 | 86.43 | |
| ODINBackbone=DenseNet, In-Distribution Dataset=CIFAR-102023.08 | 57.5 | 82.38 | |
| MSPBackbone=ResNet-34 (TIMM)2025.05 | 58.52 | 85.96 | |
| Proj. GradsBackbone=ResNet-34 (TIMM)2025.05 | 58.63 | 74.32 | |
| Softmax scoreBackbone=DenseNet, In-Distribution Dataset=CIFAR-102023.08 | 64.15 | 88.15 | |
| FedATOLalpha=0.12024.10 | 66.33 | 78.77 | |
| FedAvgalpha=0.12024.10 | 80.53 | 66.23 | |
| FedIIRalpha=0.12024.10 | 91.72 | 62.32 | |
| FedLNalpha=0.12024.10 | 93.9 | 71.99 | |
| DALBackbone=ResNet-182024.05 | 95 | 99.74 | |
| ReActBackbone=ResNetV2-50 (BiT-M)2025.05 | 95.77 | 41.93 | |
| DICEBackbone=ResNetV2-50 (BiT-M)2025.05 | 100 | 22.87 | |
| EBO-OEBackbone=ResNet-182024.05 | 111 | 99.61 | |
| DivOEBackbone=ResNet-182024.05 | 120 | 99.59 | |
| DOSBackbone=ResNet-182024.05 | 128 | 99.63 | |
| MSP-OEBackbone=ResNet-182024.05 | 229 | 99.57 | |
| DOEBackbone=ResNet-182024.05 | 275 | 99.35 | |
| MixOEBackbone=ResNet-182024.05 | 468 | 98.91 | |
| PALMMethod type=Training, Augmentations=Strong, Auxiliary outlier data=false, Backbone=ResNet-182024.05 | 1,732 | 96.82 | |
| SSD+Method type=Training, Augmentations=Strong, Auxiliary outlier data=false, Backbone=ResNet-182024.05 | 2,120 | 96.46 | |
| NPOSMethod type=Training, Augmentations=Strong, Auxiliary outlier data=false, Backbone=ResNet-182024.05 | 2,772 | 95.36 | |
| GENMethod type=Post-hoc, Augmentations=Weak, Auxiliary outlier data=false, Backbone=ResNet-182024.05 | 4,462 | 90.12 | |
| EBOMethod type=Post-hoc, Augmentations=Weak, Auxiliary outlier data=false, Backbone=ResNet-182024.05 | 4,467 | 89.61 | |
| MaxLogitMethod type=Post-hoc, Augmentations=Weak, Auxiliary outlier data=false, Backbone=ResNet-182024.05 | 4,497 | 89.56 | |
| ASHMethod type=Post-hoc, Augmentations=Weak, Auxiliary outlier data=false, Backbone=ResNet-182024.05 | 4,608 | 88.32 | |
| MSPMethod type=Post-hoc, Augmentations=Weak, Auxiliary outlier data=false, Backbone=ResNet-182024.05 | 5,504 | 90.1 |