Out-of-Distribution Detection on CIFAR-100 (ID) vs Smooth (OOD)
99.9AUCACET
Evaluation results
| Method | Links | ||||
|---|---|---|---|---|---|
| ACETAdversarial radius (epsilon)=0.012021.06 | 99.9 | 73.02 | 0 | 12.8 | |
| OEAdversarial radius (epsilon)=0.012021.06 | 99.5 | 77.25 | 0 | 0.5 | |
| ProoDAdversarial radius (epsilon)=0.01, Delta (bias shift)=52021.06 | 98.9 | 77.16 | 9 | 9 | |
| ATOMAdversarial radius (epsilon)=0.012021.06 | 98.2 | 68.32 | 0 | 80.7 | |
| DDUBackbone=ResNet-50, Single-pass capability=true2022.02 | 97.1 | — | — | — | |
| DDUBackbone=ResNet-50, Normalization=Spectral2022.02 | 97.1 | — | — | — | |
| NUQBackbone=ResNet-50, Single-pass capability=true, Density estimate=GMM2022.02 | 96.8 | — | — | — | |
| NUQBackbone=ResNet-50, Normalization=Spectral2022.02 | 96.8 | — | — | — | |
| DUQBackbone=ResNet-50, Single-pass capability=true2022.02 | 91.1 | — | — | — | |
| DUQ end-to-endBackbone=ResNet-50, Training=End-to-end2022.02 | 91.1 | — | — | — | |
| PlainAdversarial radius (epsilon)=0.012021.06 | 86.6 | 77.38 | 0 | 0.4 | |
| DUQ HeadBackbone=ResNet-50, Training=Pre-trained feature extractor2022.02 | 83.8 | — | — | — | |
| EnsembleBackbone=ResNet-50, Single-pass capability=false, Number of models=52022.02 | 83.7 | — | — | — | |
| EnsemblesBackbone=ResNet-502022.02 | 83.7 | — | — | — | |
| EntropyBackbone=ResNet-50, Single-pass capability=true2022.02 | 77.8 | — | — | — | |
| MaxProbBackbone=ResNet-50, Single-pass capability=true2022.02 | 76.6 | — | — | — | |
| TTABackbone=ResNet-50, Single-pass capability=false2022.02 | 73.2 | — | — | — | |
| TTABackbone=ResNet-50, Protocol=Test-Time Augmentation2022.02 | 73.2 | — | — | — | |
| EnergyBackbone=ResNet-50, Single-pass capability=true2022.02 | 71.5 | — | — | — | |
| DropoutBackbone=ResNet-50, Single-pass capability=false2022.02 | 63.3 | — | — | — | |
| SNGPBackbone=ResNet-50, Single-pass capability=true2022.02 | 60.9 | — | — | — | |
| ProoD-DiscAdversarial radius (epsilon)=0.012021.06 | 29.6 | — | 26.4 | 26.5 |