Out-of-Distribution Detection on ImageNet-1k vs iNaturalist
99.57AUROCNNGuide
Evaluation results
| Method | Links | ||||
|---|---|---|---|---|---|
| NNGuide2024.10 | 99.57 | 1.83 | — | — | |
| ViM2024.10 | 99.41 | 2.6 | — | — | |
| DINOv2+MLSBackbone=DINOv2, Training=Linear Probe, Postprocessor=MLS2024.10 | 98.41 | 5.64 | — | — | |
| ViT-B+CE+RMDSBackbone=ViT-B, Training=Cross Entropy (CE), Postprocessor=RMDS2024.10 | 96.1 | 19.47 | — | — | |
| RMDSBackbone=Swin-T, ID Accuracy=81.60%2026.04 | 95.57 | 18.55 | — | — | |
| OpenMaxBackbone=Swin-T, ID Accuracy=81.60%2026.04 | 95.06 | 19.56 | — | — | |
| ViMBackbone=Swin-T, ID Accuracy=81.60%2026.04 | 94.62 | 17.98 | — | — | |
| GENBackbone=Swin-T, ID Accuracy=81.60%2026.04 | 94.23 | 20.47 | — | — | |
| Energy Profile Divergence (EPD)Backbone=Swin-T, ID Accuracy=81.60%2026.04 | 94.07 | 20.37 | — | — | |
| MDSBackbone=Swin-T, ID Accuracy=81.60%2026.04 | 93.6 | 21.89 | — | — | |
| SHEBackbone=Swin-T, ID Accuracy=81.60%2026.04 | 92.78 | 33.25 | — | — | |
| TempScaleBackbone=Swin-T, ID Accuracy=81.60%2026.04 | 90.36 | 36.84 | — | — | |
| GradNormMethod Space=Gradient, Backbone=ResNetv2-101, Pre-trained on=ImageNet-1k, Input Resolution=480x4802021.10 | 90.33 | 50.03 | — | — | |
| ReActBackbone=Swin-T, ID Accuracy=81.60%2026.04 | 90.08 | 31.34 | — | — | |
| MSPBackbone=Swin-T, ID Accuracy=81.60%2026.04 | 89.84 | 37.5 | — | — | |
| ODINMethod Space=Output, Backbone=ResNetv2-101, Pre-trained on=ImageNet-1k, Input Resolution=480x4802021.10 | 89.36 | 62.69 | — | — | |
| MLSBackbone=Swin-T, ID Accuracy=81.60%2026.04 | 89.01 | 49.65 | — | — | |
| KNNBackbone=Swin-T, ID Accuracy=81.60%2026.04 | 88.91 | 31.15 | — | — | |
| EnergyMethod Space=Output, Backbone=ResNetv2-101, Pre-trained on=ImageNet-1k, Input Resolution=480x4802021.10 | 88.48 | 64.91 | — | — | |
| MSPMethod Space=Output, Backbone=ResNetv2-101, Pre-trained on=ImageNet-1k, Input Resolution=480x4802021.10 | 87.59 | 63.69 | — | — | |
| EBOBackbone=Swin-T, ID Accuracy=81.60%2026.04 | 85.17 | 61.18 | — | — | |
| DICEBackbone=Swin-T, ID Accuracy=81.60%2026.04 | 47.09 | 98.11 | — | — | |
| ASHBackbone=Swin-T, ID Accuracy=81.60%2026.04 | 46.49 | 94.19 | — | — | |
| MahalanobisMethod Space=Feature, Backbone=ResNetv2-101, Pre-trained on=ImageNet-1k, Input Resolution=480x4802021.10 | 46.33 | 96.34 | — | — | |
| GradNormBackbone=Swin-T, ID Accuracy=81.60%2026.04 | 38.28 | 95.18 | — | — |