Out-of-distribution Detection on CIFAR-10 vs CIFAR-100 (test)
97.63AUROCForte+GMM
Evaluation results
| Method | Links | |||||||||||||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Forte+GMMSupervision=Unsupervised2024.10 | 97.63 | 9.69 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| Forte+SVMSupervision=Unsupervised2024.10 | 97.29 | 9.08 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| Forte+KDESupervision=Unsupervised2024.10 | 94.81 | 18.2 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| SSDk+k=52021.03 | 94.1 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| SSD+2021.03 | 93.4 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| Outlier exposureapproach_type=Classifier-based2019.09 | 93.3 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| Outlier exposureBackbone=Wide ResNet-50, Uses labeled OOD data for training=true2020.07 | 93.3 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| Outlier exposure2021.03 | 93.3 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| Contrastive Training for OOD DetectionBackbone=Wide ResNet-502020.07 | 92.9 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| Contrastive + SupervisedBackbone=4x wider ResNet-50 network2021.03 | 92.9 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| AHGCBackbone=ResNet-18, Optimizer=SGD2024.12 | 92.72 | 24.82 | — | — | — | — | — | — | — | — | — | — | — | 93.5 | 91.69 | 3.81 | 21 | 47.87 | 80.04 | |
| CSI2021.03 | 92.2 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| SconeBackbone=ResNet-18, Optimizer=SGD2024.12 | 91.24 | 43.79 | — | — | — | — | — | — | — | — | — | — | — | 90.68 | 88.76 | 2.38 | 8.17 | 42.96 | 77.94 | |
| UDGBackbone=ResNet-18, Optimizer=SGD2024.12 | 90.98 | 47.2 | — | — | — | — | — | — | — | — | — | — | — | 91.74 | 89.36 | 1.5 | 10.94 | 40.34 | 75.89 | |
| Auxiliary Rotation Predictionaverage_of_runs=52019.06 | 90.9 | 44.9 | 67.7 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| Rotation pred.Backbone=Wide ResNet-50, Uses additional data for pretraining=true2020.07 | 90.9 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| Rotation-loss + Supervised2021.03 | 90.9 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| MahalanobisRequires additional out-of-distribution data for tuning=true2021.03 | 90.5 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| Softmax-probs2021.03 | 89.8 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| SNN2026.01 | 89.8 | 50.1 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| ODINRequires additional out-of-distribution data for tuning=true2021.03 | 89.6 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| Comp.VAEvariant=ours2026.01 | 89.5 | 23.2 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| Residual flowsBackbone=Wide ResNet-50, Uses labeled OOD data for tuning=true2020.07 | 89.4 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| Residual FlowsRequires additional out-of-distribution data for tuning=true2021.03 | 89.4 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| SupConScore=max_y p(y|x)2023.03 | 89 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| ReAct2026.01 | 88.96 | 53.51 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| ODIN2026.01 | 88.9 | 52.32 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| MahalanobisBackbone=Wide ResNet-50, Uses labeled OOD data for tuning=true2020.07 | 88.2 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| LogitNorm2026.01 | 88.03 | 55.08 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| VIBScore=max_y p(y|x)2023.03 | 88 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| VIBScore=log Dir_0.05(y)2023.03 | 88 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| TSObjective Category=Baseline2026.06 | 88 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| KNN2026.01 | 87.9 | 58.34 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| CEObjective Category=Baseline2026.06 | 87.82 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| ConjNormBackbone=ResNet-18, Optimizer=SGD2024.12 | 87.46 | 30.94 | — | — | — | — | — | — | — | — | — | — | — | 91.18 | 89.23 | 1.06 | 13.78 | 39.87 | 75.5 | |
| AdaFocalObjective Category=Baseline2026.06 | 87.3 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| Maximum Softmax Probability (MSP)average_of_runs=52019.06 | 87.2 | 45.9 | 54.1 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| DICE2026.01 | 87.11 | 58.6 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| JEMtype=generative-based, variant=best results2021.04 | 87 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| JEMScore=max_y p(y|x)2023.03 | 87 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| DFLObjective Category=Baseline2026.06 | 86.43 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| Softmax probs.Backbone=Wide ResNet-502020.07 | 86.4 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| OEBackbone=ResNet-18, Optimizer=SGD2024.12 | 86.22 | 58.54 | — | — | — | — | — | — | — | — | — | — | — | 86.17 | 84.88 | 3.64 | 6.55 | 19.04 | 61.11 | |
| Energy score2026.01 | 86.06 | 58.66 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| Cross-entropyScore=max_y p(y|x)2023.03 | 86 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| Cross-entropyScore=log Dir_0.05(y)2023.03 | 86 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| AENIBScore=log Dir_0.05(y)2023.03 | 86 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| AENIBScore=log Dir_0.05(y) + log N(z_n; 0, I)2023.03 | 86 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| ODINBackbone=Wide ResNet-50, Uses labeled OOD data for tuning=true2020.07 | 85.8 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| MSP2026.01 | 85.28 | 64.66 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| ACLSObjective Category=Baseline2026.06 | 84.25 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| FeDLaS-MbLSObjective Category=MbLS-based2026.06 | 84.15 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| SSDk+k=52021.03 | 84.1 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| MOODFLOPs=0.79 x 10^82021.04 | 84 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| AENIBScore=max_y p(y|x)2023.03 | 84 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| MCDBackbone=ResNet-18, Optimizer=SGD2024.12 | 82.78 | 74 | — | — | — | — | — | — | — | — | — | — | — | 83.97 | 79.16 | 0.8 | 4.99 | 18.88 | 58.18 | |
| GODIN2026.01 | 82.37 | 60.69 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| EBOBackbone=ResNet-18, Optimizer=SGD2024.12 | 79.65 | 56.98 | — | — | — | — | — | — | — | — | — | — | — | 75.09 | 81.23 | 0.1 | 0.69 | 4.74 | 34.28 | |
| GradNorm2026.01 | 79.31 | 65.44 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| Gram matrixBackbone=Wide ResNet-502020.07 | 79 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| Gram Matrix2021.03 | 79 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| Mahalanobis2026.01 | 78.93 | 87.71 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| ODINBackbone=ResNet-18, Optimizer=SGD2024.12 | 78.4 | 56.34 | — | — | — | — | — | — | — | — | — | — | — | 73.21 | 80.99 | 0.1 | 0.38 | 4.43 | 30.11 | |
| Contrastive Training for OOD DetectionBackbone=Wide ResNet-502020.07 | 78.3 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| Contrastive + SupervisedBackbone=4x wider ResNet-50 network2021.03 | 78.3 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| SSD+2021.03 | 78.3 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| Softmax-probs2021.03 | 78 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| ODINRequires additional out-of-distribution data for tuning=true2021.03 | 77.9 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| DFLObjective Category=Baseline2026.06 | 77.6 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| MahalanobisBackbone=Wide ResNet-50, Uses labeled OOD data for tuning=true2020.07 | 77.5 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| Comp.VAEdistribution=vMF2026.01 | 77.4 | 61.1 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| ODINBackbone=Wide ResNet-50, Uses labeled OOD data for tuning=true2020.07 | 77.2 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| FeDLaS-LSObjective Category=LS-based2026.06 | 77.2 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| MbLSObjective Category=MbLS-based2026.06 | 77.2 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| Softmax probs.Backbone=Wide ResNet-502020.07 | 77.1 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| Residual flowsBackbone=Wide ResNet-50, Uses labeled OOD data for tuning=true2020.07 | 77.1 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| Residual FlowsRequires additional out-of-distribution data for tuning=true2021.03 | 77.1 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| AdaFocalObjective Category=Baseline2026.06 | 76.87 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| Outlier exposureBackbone=Wide ResNet-50, Uses labeled OOD data for training=true2020.07 | 75.7 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| Outlier exposure2021.03 | 75.7 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| LSObjective Category=LS-based2026.06 | 75.2 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| TSObjective Category=Baseline2026.06 | 74.84 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| CEObjective Category=Baseline2026.06 | 74.82 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| Serrà et altype=likelihood ratio-based, variant=best results, FLOPs=2.78 x 10^102021.04 | 74 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| FeDLaS-MbLSObjective Category=MbLS-based2026.06 | 73.72 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| FeDLaS-LSObjective Category=LS-based2026.06 | 73.71 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| S scoregenerative_model=Glow, compression_algorithm=FLIF2019.09 | 73.6 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| MbLSObjective Category=MbLS-based2026.06 | 73.19 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| LSObjective Category=LS-based2026.06 | 72.33 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| ACLSObjective Category=Baseline2026.06 | 71.76 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| Outlier exposureapproach_type=Generative-based2019.09 | 68.5 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| Gram matrixBackbone=Wide ResNet-502020.07 | 67.9 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| Gram Matrix2021.03 | 67.9 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| JEMScore=log p(x)2023.03 | 67 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| Glowtype=generative-based, FLOPs=4.09 x 10^92021.04 | 65 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| KNN*2026.01 | 61.5 | 90 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| DOSESupervision=Unsupervised2024.10 | 56.9 | 91.95 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| MahalanobisRequires additional out-of-distribution data for tuning=true2021.03 | 55.3 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| TTSupervision=Unsupervised2024.10 | 54.8 | 92.06 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| Sketched Lanczos UncertaintyMemory budget=10p2024.09 | 54 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — |