Out-of-distribution Detection on CIFAR-100 vs SVHN (test)
99.1AUROCSSDk
Evaluation results
| Method | Links | |||||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|
| SSDkk=52021.03 | 99.1 | — | — | — | — | — | — | — | — | — | — | |
| Reg. MahalanobisBackbone=ResNet182026.02 | 98.78 | — | — | — | — | — | 96.97 | — | — | — | — | |
| LMD2026.05 | 98.5 | — | — | — | — | — | — | — | — | — | — | |
| Mahalanobis distance-based scoreBackbone=ResNet, Validation Setup=OOD samples2018.07 | 98.4 | 91.9 | 93.7 | — | — | — | — | — | — | — | — | |
| MahalanobisBackbone=ResNet2018.07 | 98.4 | 91.9 | 93.7 | 96.4 | 99.3 | — | — | — | — | — | — | |
| MahalanobisBackbone=ResNet2019.12 | 98.4 | 91.9 | 93.7 | — | — | — | — | — | — | — | — | |
| MahalanobisModel Architecture=ResNet2019.12 | 98.4 | 91.9 | 93.7 | — | — | — | — | — | — | — | — | |
| COMBOODBackbone=ResNet182026.02 | 97.69 | — | — | — | — | — | 95.2 | — | — | — | — | |
| OursBackbone=DenseNet2019.12 | 97.3 | 89.3 | 92.4 | — | — | — | — | — | — | — | — | |
| OursModel Architecture=DenseNet2019.12 | 97.3 | 89.3 | 92.4 | — | — | — | — | — | — | — | — | |
| Mahalanobis distance-based scoreBackbone=DenseNet, Validation Setup=OOD samples2018.07 | 97.2 | 82.5 | 91.5 | — | — | — | — | — | — | — | — | |
| MahalanobisBackbone=DenseNet2018.07 | 97.2 | 82.5 | 91.5 | 94.8 | 98.8 | — | — | — | — | — | — | |
| MahalanobisBackbone=DenseNet2019.12 | 97.2 | 82.5 | 91.5 | — | — | — | — | — | — | — | — | |
| MahalanobisModel Architecture=DenseNet2019.12 | 97.2 | 82.5 | 91.5 | — | — | — | — | — | — | — | — | |
| OursBackbone=ResNet2019.12 | 96 | 80.8 | 89.6 | — | — | — | — | — | — | — | — | |
| OursModel Architecture=ResNet2019.12 | 96 | 80.8 | 89.6 | — | — | — | — | — | — | — | — | |
| MDSBackbone=ResNet182026.02 | 95.68 | — | — | — | — | — | 90.63 | — | — | — | — | |
| lambda (Similarity-only)2026.02 | 95.3 | — | — | — | — | — | — | — | — | — | — | |
| SSD2021.03 | 94.9 | — | — | — | — | — | — | — | — | — | — | |
| SPEM-noiseFlow Model=ResFlow, Feature Extractor=ResNet-152, Hyperparameter alpha=0.12026.02 | 94.58 | — | — | — | — | — | — | — | — | — | — | |
| DM Dual Threshold2026.05 | 94.5 | — | — | — | — | — | — | — | — | — | — | |
| Proj. Regret2026.05 | 94.5 | — | — | — | — | — | — | — | — | — | — | |
| Rotation-loss2021.03 | 94.4 | — | — | — | — | — | — | — | — | — | — | |
| ODINBackbone=ResNet, Validation Setup=OOD samples2018.07 | 93.9 | 62.7 | 88 | — | — | — | — | — | — | — | — | |
| ODINBackbone=ResNet2018.07 | 93.9 | 62.7 | 88 | 89 | 96.9 | — | — | — | — | — | — | |
| ODINBackbone=ResNet2019.12 | 93.9 | 62.7 | 88 | — | — | — | — | — | — | — | — | |
| ODINModel Architecture=ResNet2019.12 | 93.9 | 62.7 | 88 | — | — | — | — | — | — | — | — | |
| ODINBackbone=DenseNet, Validation Setup=OOD samples2018.07 | 93.8 | 70.6 | 86.6 | — | — | — | — | — | — | — | — | |
| ODINBackbone=DenseNet2018.07 | 93.8 | 70.6 | 86.6 | 87.1 | 97.3 | — | — | — | — | — | — | |
| ODINBackbone=DenseNet2019.12 | 93.8 | 70.6 | 86.6 | — | — | — | — | — | — | — | — | |
| ODINModel Architecture=DenseNet2019.12 | 93.8 | 70.6 | 86.6 | — | — | — | — | — | — | — | — | |
| SPEMFlow Model=ResFlow, Feature Extractor=ResNet-152, Hyperparameter alpha=0.42026.02 | 93.59 | — | — | — | — | — | — | — | — | — | — | |
| LIDFlow Model=ResFlow2026.02 | 93.3 | — | — | — | — | — | — | — | — | — | — | |
| Mahalanobis distance-based scoreBackbone=DenseNet, Validation Setup=adversarial samples2018.07 | 91.8 | 62.2 | 84.6 | — | — | — | — | — | — | — | — | |
| E1 normed-forkBackbone=CLIP, Embedding Dimension=512-d2026.05 | 90.3 | — | — | — | — | 50.2 | — | — | — | — | — | |
| REGradBackbone=ViT2024.04 | 90.22 | — | — | — | — | — | 88.61 | — | — | — | — | |
| NUQBackbone=ResNet-50, Single-pass capability=true, Density estimate=GMM2022.02 | 89.7 | — | — | — | — | — | — | — | — | — | — | |
| DDUBackbone=ResNet-50, Single-pass capability=true2022.02 | 89.6 | — | — | — | — | — | — | — | — | — | — | |
| Clipped HNNScoring=Softmax scores, In-distribution dataset=CIFAR1002021.07 | 89.53 | — | — | — | — | 53.11 | 97.71 | — | — | — | — | |
| DUQBackbone=ResNet-50, Single-pass capability=true2022.02 | 88.7 | — | — | — | — | — | — | — | — | — | — | |
| MahalanobisBackbone=ResNet-502026.05 | 88.4 | — | — | — | — | 47.5 | — | — | — | — | — | |
| ODINBackbone=DenseNet, Validation Setup=adversarial samples2018.07 | 88.2 | 39.8 | 80.7 | — | — | — | — | — | — | — | — | |
| k-NNBackbone=DINOv2, k-value=502026.05 | 88.1 | — | — | — | — | 66.5 | — | — | — | — | — | |
| REGrad*layer-selective=true, perturbation=true2024.04 | 88.06 | — | — | — | — | — | 79.11 | — | — | — | — | |
| Typicality (Entropy)Flow Model=ResFlow2026.02 | 87.83 | — | — | — | — | — | — | — | — | — | — | |
| MedixBackbone=ResNet-34, Wild OOD data (Pout)=300K Random Images2025.10 | 87.25 | — | — | — | — | 41.29 | — | — | — | — | — | |
| Outlier ExposureBackbone=Wide ResNet 40x22021.05 | 86.9 | — | — | — | — | 42.9 | — | — | — | — | — | |
| Deep Abstaining ClassifierBackbone=ResNet 342021.05 | 86.85 | — | — | — | — | 41.31 | — | — | — | — | — | |
| WOODSBackbone=ResNet-34, Wild OOD data (Pout)=300K Random Images2025.10 | 86.76 | — | — | — | — | 69.41 | — | — | — | — | — | |
| Perturb x2024.04 | 86.49 | — | — | — | — | — | 86.76 | — | — | — | — | |
| SNGPBackbone=ResNet-50, Single-pass capability=true2022.02 | 86.2 | — | — | — | — | — | — | — | — | — | — | |
| Energy (w/ OE)Backbone=ResNet-34, Wild OOD data (Pout)=300K Random Images2025.10 | 85.59 | — | — | — | — | 69.81 | — | — | — | — | — | |
| ExGradBackbone=ViT2024.04 | 85.47 | — | — | — | — | — | 78.91 | — | — | — | — | |
| GMMFlow Model=ResFlow2026.02 | 85.46 | — | — | — | — | — | — | — | — | — | — | |
| Deep Abstaining ClassifierBackbone=Wide ResNet 40x22021.05 | 85.44 | — | — | — | — | 40.46 | — | — | — | — | — | |
| EnergyBackbone=ResNet-502026.05 | 85.2 | — | — | — | — | 74.1 | — | — | — | — | — | |
| Mahalanobis distance-based scoreBackbone=ResNet, Validation Setup=adversarial samples2018.07 | 84.4 | 41.9 | 76.5 | — | — | — | — | — | — | — | — | |
| Deep Mahalanobis DetectorBackbone=ResNet 342021.05 | 84.4 | — | — | — | — | 58.1 | — | — | — | — | — | |
| Perturb θBackbone=ViT2024.04 | 84.35 | — | — | — | — | — | 77.37 | — | — | — | — | |
| ENNScoring=Softmax scores, In-distribution dataset=CIFAR1002021.07 | 84.32 | — | — | — | — | 84.56 | 96.69 | — | — | — | — | |
| KNNBackbone=ResNet182026.02 | 84.15 | — | — | — | — | — | 92.79 | — | — | — | — | |
| EnsembleBackbone=ResNet-50, Single-pass capability=false, Number of models=52022.02 | 82.9 | — | — | — | — | — | — | — | — | — | — | |
| OEBackbone=ResNet-34, Wild OOD data (Pout)=300K Random Images2025.10 | 82.89 | — | — | — | — | 68.8 | — | — | — | — | — | |
| BaselineBackbone=DenseNet, Validation Setup=OOD samples2018.07 | 82.7 | 26.7 | 75.6 | — | — | — | — | — | — | — | — | |
| BaselineBackbone=DenseNet, Validation Setup=adversarial samples2018.07 | 82.7 | 26.7 | 75.6 | — | — | — | — | — | — | — | — | |
| BaselineBackbone=DenseNet2018.07 | 82.7 | 26.7 | 75.6 | 74.3 | 91 | — | — | — | — | — | — | |
| BaselineBackbone=DenseNet2019.12 | 82.7 | 26.7 | 75.6 | — | — | — | — | — | — | — | — | |
| BaselineModel Architecture=DenseNet2019.12 | 82.7 | 26.7 | 75.6 | — | — | — | — | — | — | — | — | |
| EnergyBackbone=ResNet-50, Single-pass capability=true2022.02 | 82 | — | — | — | — | — | — | — | — | — | — | |
| TTABackbone=ResNet-50, Single-pass capability=false2022.02 | 81.6 | — | — | — | — | — | — | — | — | — | — | |
| EntropyBackbone=ResNet-50, Single-pass capability=true2022.02 | 81.1 | — | — | — | — | — | — | — | — | — | — | |
| OpenMaxBackbone=ResNet 342021.05 | 80.88 | — | — | — | — | 53.22 | — | — | — | — | — | |
| EncMin2L2026.05 | 80.6 | — | — | — | — | 64.8 | — | — | — | — | — | |
| LA2024.04 | 80.09 | — | — | — | — | — | 73.63 | — | — | — | — | |
| Exgrad2024.04 | 79.99 | — | — | — | — | — | 72.68 | — | — | — | — | |
| MaxProbBackbone=ResNet-50, Single-pass capability=true2022.02 | 79.7 | — | — | — | — | — | — | — | — | — | — | |
| Likelihood RatioFlow Model=ResFlow2026.02 | 79.52 | — | — | — | — | — | — | — | — | — | — | |
| BaselineBackbone=ResNet, Validation Setup=OOD samples2018.07 | 79.5 | 20.3 | 73.2 | — | — | — | — | — | — | — | — | |
| BaselineBackbone=ResNet, Validation Setup=adversarial samples2018.07 | 79.5 | 20.3 | 73.2 | — | — | — | — | — | — | — | — | |
| BaselineBackbone=ResNet2018.07 | 79.5 | 20.3 | 73.2 | 64.8 | 89 | — | — | — | — | — | — | |
| BaselineBackbone=ResNet2019.12 | 79.5 | 20.3 | 73.2 | — | — | — | — | — | — | — | — | |
| BaselineModel Architecture=ResNet2019.12 | 79.5 | 20.3 | 73.2 | — | — | — | — | — | — | — | — | |
| MSPBackbone=ResNet-502026.05 | 78.4 | — | — | — | — | 78.6 | — | — | — | — | — | |
| DropoutBackbone=ResNet-50, Single-pass capability=false2022.02 | 77.6 | — | — | — | — | — | — | — | — | — | — | |
| k-NNBackbone=ResNet-50, k-value=502026.05 | 77 | — | — | — | — | 89.2 | — | — | — | — | — | |
| VDM log pT2026.05 | 75.5 | — | — | — | — | 63.3 | — | — | — | — | — | |
| Score NetworkConfig=Hk = 4dk2026.05 | 74.4 | — | — | — | — | 75.8 | — | — | — | — | — | |
| k-NNBackbone=CLIP, k-value=502026.05 | 73.4 | — | — | — | — | 81.8 | — | — | — | — | — | |
| ComplexityFlow Model=ResFlow2026.02 | 73.31 | — | — | — | — | — | — | — | — | — | — | |
| DiffPath-6D2026.05 | 72.4 | — | — | — | — | — | — | — | — | — | — | |
| ODINBackbone=ResNet, Validation Setup=adversarial samples2018.07 | 72 | 12.2 | 67.7 | — | — | — | — | — | — | — | — | |
| ODINBackbone=ResNet182026.02 | 71.08 | — | — | — | — | — | 52.36 | — | — | — | — | |
| UNGradBackbone=ViT2024.04 | 70.2 | — | — | — | — | — | 60.06 | — | — | — | — | |
| EigenScore2026.05 | 66.1 | — | — | — | — | — | — | — | — | — | — | |
| E2 normed-forkBackbone=DINOv2, Embedding Dimension=768-d2026.05 | 64.2 | — | — | — | — | 75 | — | — | — | — | — | |
| Inserted Dropout2024.04 | 57.39 | — | — | — | — | — | 55.49 | — | — | — | — | |
| GramBackbone=ResNet182026.02 | 56.69 | — | — | — | — | — | 28.53 | — | — | — | — | |
| GradNormBackbone=ViT2024.04 | 56.3 | — | — | — | — | — | 54.84 | — | — | — | — | |
| MC-AA2024.04 | 54.74 | — | — | — | — | — | 59.26 | — | — | — | — | |
| Entropy2024.04 | 53.38 | — | — | — | — | — | 55.2 | — | — | — | — |