OOD Detection on ImageNet Average
20.7FPR95LINe
Evaluation results
| Method | Links | |||
|---|---|---|---|---|
| LINePost-hoc=true2023.03 | 20.7 | 95.03 | — | |
| ASH-BBackbone=ResNet, In-Distribution Dataset=ImageNet-1k2022.09 | 22.73 | 95.06 | — | |
| ASH-SBackbone=ResNet, In-Distribution Dataset=ImageNet-1k2022.09 | 22.8 | 95.12 | — | |
| DICE + ReActPost-hoc=true2023.03 | 27.25 | 93.4 | — | |
| DICE + ReActBackbone=ResNet, In-Distribution Dataset=ImageNet-1k2022.09 | 27.25 | 93.4 | — | |
| ReActPost-hoc=true2023.03 | 31.43 | 92.95 | — | |
| ReActBackbone=ResNet, In-Distribution Dataset=ImageNet-1k2022.09 | 31.43 | 92.95 | — | |
| DICE + ReActBackbone=MobileNet, In-Distribution Dataset=ImageNet-1k2022.09 | 31.64 | 92.68 | — | |
| DICEPost-hoc=true2023.03 | 34.75 | 90.77 | — | |
| DICEBackbone=ResNet, In-Distribution Dataset=ImageNet-1k2022.09 | 34.75 | 90.77 | — | |
| ASH-BBackbone=MobileNet, In-Distribution Dataset=ImageNet-1k2022.09 | 35.66 | 92.13 | — | |
| ASH-SBackbone=MobileNet, In-Distribution Dataset=ImageNet-1k2022.09 | 38.67 | 90.95 | — | |
| DICEBackbone=MobileNet, In-Distribution Dataset=ImageNet-1k2022.09 | 41.92 | 89.6 | — | |
| ReActBackbone=MobileNet, In-Distribution Dataset=ImageNet-1k2022.09 | 45.02 | 89.47 | — | |
| ASH-PBackbone=ResNet, In-Distribution Dataset=ImageNet-1k2022.09 | 50.32 | 89.04 | — | |
| diverseMixBackbone=ResNet2024.11 | 53.07 | 89.36 | 92.13 | |
| ODINBackbone=MobileNet, In-Distribution Dataset=ImageNet-1k2022.09 | 54.2 | 85.81 | — | |
| ODINPost-hoc=true2023.03 | 56.48 | 85.41 | — | |
| ODINBackbone=ResNet, In-Distribution Dataset=ImageNet-1k2022.09 | 56.48 | 85.41 | — | |
| OEBackbone=ResNet2024.11 | 56.61 | 89.05 | 91.79 | |
| ASH-PBackbone=MobileNet, In-Distribution Dataset=ImageNet-1k2022.09 | 57.15 | 87.34 | — | |
| MixOEBackbone=ResNet2024.11 | 57.68 | 87.38 | 90.89 | |
| ODINBackbone=ResNet2024.11 | 58.16 | 86.82 | 91 | |
| EnergyPost-hoc=true2023.03 | 58.41 | 86.17 | — | |
| Energy scoreBackbone=ResNet, In-Distribution Dataset=ImageNet-1k2022.09 | 58.41 | 86.17 | — | |
| Energy (w. Daux)Backbone=ResNet2024.11 | 59.11 | 88.22 | 90.5 | |
| Max LogitsBackbone=ResNet2024.11 | 59.28 | 86.66 | 90.75 | |
| EnergyBackbone=ResNet2024.11 | 60.46 | 86.33 | 90.38 | |
| MSPBackbone=ResNet2024.11 | 60.85 | 86.39 | 90.54 | |
| DPNBackbone=ResNet2024.11 | 62.14 | 87.89 | 90.68 | |
| Energy scoreBackbone=MobileNet, In-Distribution Dataset=ImageNet-1k2022.09 | 62.39 | 84.91 | — | |
| MahalanobisBackbone=MobileNet, In-Distribution Dataset=ImageNet-1k2022.09 | 63.6 | 71.01 | — | |
| MSPPost-hoc=true2023.03 | 66.95 | 81.99 | — | |
| Softmax scoreBackbone=ResNet, In-Distribution Dataset=ImageNet-1k2022.09 | 66.95 | 81.99 | — | |
| Softmax scoreBackbone=MobileNet, In-Distribution Dataset=ImageNet-1k2022.09 | 73.51 | 79 | — | |
| MahalanobisPost-hoc=true2023.03 | 87.43 | 55.47 | — | |
| MahalanobisBackbone=ResNet, In-Distribution Dataset=ImageNet-1k2022.09 | 87.43 | 55.47 | — |