Out-of-distribution Detection on ImageNet OOD Average 1k (test)
16.53FPR@95NNGuide
Evaluation results
| Method | Links | |||
|---|---|---|---|---|
| NNGuideTraining scheme=Transfer learning, Model=RegNet-Y/16GF, ID accuracy=86.01%2023.09 | 16.53 | 95.89 | 98.98 | |
| GradOrthBackbone=ResNet2023.08 | 18.57 | 96.31 | — | |
| VIMTraining scheme=Transfer learning, Model=RegNet-Y/16GF, ID accuracy=86.01%2023.09 | 21.39 | 94.89 | 98.74 | |
| TV-OODBackbone=ViT-B_16, Auxiliary OOD dataset=with auxiliary2026.01 | 22.28 | 94.81 | 98.77 | |
| ASH-BBackbone=ResNet2023.08 | 22.73 | 95.06 | — | |
| ASH-SBackbone=ResNet2023.08 | 22.8 | 95.12 | — | |
| WOODSBackbone=ViT-B_16, Auxiliary OOD dataset=with auxiliary2026.01 | 22.95 | 93.38 | 98.56 | |
| OEBackbone=ViT-B_16, Auxiliary OOD dataset=with auxiliary2026.01 | 22.97 | 95 | 81.35 | |
| EnergyTraining scheme=Transfer learning, Model=RegNet-Y/16GF, ID accuracy=86.01%2023.09 | 24.73 | 93.08 | 98.18 | |
| KLTraining scheme=Transfer learning, Model=RegNet-Y/16GF, ID accuracy=86.01%2023.09 | 24.74 | 93.07 | 98.18 | |
| BEOEBackbone=ViT-B_16, Auxiliary OOD dataset=with auxiliary2026.01 | 25.31 | 93.86 | 98.66 | |
| VRA-PBackbone=ResNet2023.08 | 25.49 | 94.57 | — | |
| MaxLogitTraining scheme=Transfer learning, Model=RegNet-Y/16GF, ID accuracy=86.01%2023.09 | 26 | 92.91 | 98.16 | |
| EFBackbone=ViT-B_16, Auxiliary OOD dataset=with auxiliary2026.01 | 26.58 | 92.69 | 98.22 | |
| DICE + ReActBackbone=ResNet-50, Pre-trained=ImageNet-1k2021.11 | 27.25 | 93.4 | — | |
| DICE + ReActBackbone=ResNet2023.08 | 27.25 | 93.4 | — | |
| GradOrthBackbone=MobileNet2023.08 | 27.65 | 93.25 | — | |
| NNGuideTraining scheme=From scratch, Model=ResNet-50, ID accuracy=78.73%2023.09 | 27.81 | 92.89 | 98.03 | |
| ZODE-KNN2022.12 | 28.1 | 93.99 | — | |
| LogitNormBackbone=ViT-B_16, Auxiliary OOD dataset=w/o auxiliary2026.01 | 28.41 | 93.39 | 98.69 | |
| VRA-DNBackbone=ResNet2023.08 | 28.53 | 94.09 | — | |
| GradNormTraining scheme=From scratch, Model=ResNet-50, ID accuracy=78.73%2023.09 | 28.92 | 93 | 98.09 | |
| KNNBackbone=ViT-B_16, Auxiliary OOD dataset=w/o auxiliary2026.01 | 29.31 | 92.58 | 98.59 | |
| VIMTraining scheme=From scratch, Model=ResNet-50, ID accuracy=78.73%2023.09 | 29.86 | 93 | 98.13 | |
| TV-OODBackbone=ViT-B_16, Auxiliary OOD dataset=w/o auxiliary2026.01 | 30.01 | 92.34 | 98.75 | |
| ASH-S@90Model=DenseNet-1212022.09 | 30.3 | 93.09 | — | |
| ASH-B@90Model=DenseNet-1212022.09 | 30.38 | 93.34 | — | |
| VIMBackbone=ViT-B_16, Auxiliary OOD dataset=w/o auxiliary2026.01 | 31.13 | 93.07 | 98.57 | |
| KNNTraining scheme=Transfer learning, Model=RegNet-Y/16GF, ID accuracy=86.01%2023.09 | 31.26 | 91.96 | 97.91 | |
| ReActBackbone=ResNet-502021.11 | 31.43 | 92.95 | — | |
| ReActBackbone=ResNet-50, Pre-trained=ImageNet-1k2021.11 | 31.43 | 92.95 | — | |
| ReActBackbone=ResNet2023.08 | 31.43 | 92.95 | — | |
| DICE + ReActBackbone=MobileNet2023.08 | 31.64 | 92.68 | — | |
| MahalanobisTraining scheme=Transfer learning, Model=RegNet-Y/16GF, ID accuracy=86.01%2023.09 | 32.15 | 93.07 | 98.36 | |
| ASH-B@90Methods Space=Feature, Backbone=Google BiT-S2023.11 | 32.18 | 93.43 | — | |
| ASH-B@65Model=VGG-162022.09 | 32.2 | 92.19 | — | |
| EnergyBackbone=ViT-B_16, Auxiliary OOD dataset=w/o auxiliary2026.01 | 32.31 | 93 | 98.52 | |
| ASH-SBackbone=ViT-B_16, Auxiliary OOD dataset=w/o auxiliary2026.01 | 34.17 | 82.35 | 98.35 | |
| NNGuideTraining scheme=Transfer learning, Model=ViT-B/16, ID accuracy=85.3%2023.09 | 34.2 | 92.14 | 98.1 | |
| ODINBackbone=ViT-B_16, Auxiliary OOD dataset=w/o auxiliary2026.01 | 34.45 | 92.58 | 98.44 | |
| DICEBackbone=ResNet-50, Pre-trained=ImageNet-1k2021.11 | 34.75 | 90.77 | — | |
| DICEBackbone=ResNet2023.08 | 34.75 | 90.77 | — | |
| VIMTraining scheme=Transfer learning, Model=ViT-B/16, ID accuracy=85.3%2023.09 | 35.16 | 91.04 | 97.73 | |
| ASH-BBackbone=MobileNet2023.08 | 35.66 | 92.13 | — | |
| RankfeatMethods Space=Feature, Backbone=Google BiT-S, Blocks=Block3+42023.11 | 36.8 | 92.15 | — | |
| GAIA-AMethods Space=Gradient, Backbone=Google BiT-S2023.11 | 37.42 | 91.9 | — | |
| MahalanobisBackbone=ViT-B_16, Auxiliary OOD dataset=w/o auxiliary2026.01 | 37.5 | 91.28 | 98.12 | |
| MaxLogitTraining scheme=Transfer learning, Model=ViT-B/16, ID accuracy=85.3%2023.09 | 37.62 | 89.29 | 97.11 | |
| GradNormTraining scheme=Transfer learning, Model=ViT-B/16, ID accuracy=85.3%2023.09 | 38.04 | 89.28 | 96.91 | |
| KLTraining scheme=Transfer learning, Model=ViT-B/16, ID accuracy=85.3%2023.09 | 38.44 | 89.01 | 97.03 | |
| EnergyTraining scheme=Transfer learning, Model=ViT-B/16, ID accuracy=85.3%2023.09 | 38.44 | 89.01 | 97.03 | |
| KNN+2022.12 | 38.47 | 90.91 | — | |
| ASH-SBackbone=MobileNet2023.08 | 38.67 | 90.95 | — | |
| MahalanobisTraining scheme=Transfer learning, Model=ViT-B/16, ID accuracy=85.3%2023.09 | 39.34 | 91.73 | 98.08 | |
| RankfeatMethods Space=Feature, Backbone=Google BiT-S, Blocks=Block42023.11 | 39.69 | 87.84 | — | |
| KLTraining scheme=From scratch, Model=ResNet-50, ID accuracy=78.73%2023.09 | 40.01 | 90.8 | 97.63 | |
| EnergyTraining scheme=From scratch, Model=ResNet-50, ID accuracy=78.73%2023.09 | 40.01 | 90.8 | 97.63 | |
| GradNormBackbone=ResNet2023.08 | 40.29 | 87.34 | — | |
| SSDTraining scheme=Transfer learning, Model=RegNet-Y/16GF, ID accuracy=86.01%2023.09 | 40.64 | 90.22 | 97.57 | |
| SSDTraining scheme=From scratch, Model=ResNet-50, ID accuracy=78.73%2023.09 | 40.94 | 91.47 | 97.82 | |
| ASH-S@99Model=VGG-162022.09 | 41.49 | 89.38 | — | |
| DICEBackbone=MobileNet2023.08 | 41.92 | 89.6 | — | |
| MaxLogitTraining scheme=From scratch, Model=ResNet-50, ID accuracy=78.73%2023.09 | 42.12 | 90.49 | 97.57 | |
| KNNTraining scheme=From scratch, Model=ResNet-50, ID accuracy=78.73%2023.09 | 42.73 | 90.19 | 97.44 | |
| MSPTraining scheme=Transfer learning, Model=RegNet-Y/16GF, ID accuracy=86.01%2023.09 | 43.37 | 88.95 | 97.3 | |
| MahalanobisTraining scheme=From scratch, Model=ResNet-50, ID accuracy=78.73%2023.09 | 44.58 | 90.93 | 97.76 | |
| ASH-S@95Model=VGG-162022.09 | 44.87 | 90.47 | — | |
| ReActBackbone=MobileNet2021.11 | 45.02 | 89.47 | — | |
| ReActBackbone=MobileNet2023.08 | 45.02 | 89.47 | — | |
| ASH-B@65Model=DenseNet-1212022.09 | 47.08 | 87.35 | — | |
| ExGradBackbone=ResNet2023.08 | 47.4 | 75.07 | — | |
| MSPBackbone=ViT-B_16, Auxiliary OOD dataset=w/o auxiliary2026.01 | 48.09 | 87.41 | 97.29 | |
| MSPTraining scheme=Transfer learning, Model=ViT-B/16, ID accuracy=85.3%2023.09 | 48.74 | 87.67 | 96.94 | |
| MSPTraining scheme=From scratch, Model=ResNet-50, ID accuracy=78.73%2023.09 | 49.54 | 87.44 | 96.71 | |
| ASH-PBackbone=ResNet2023.08 | 50.32 | 89.04 | — | |
| Energy scoreModel=DenseNet-1212022.09 | 50.4 | 87.66 | — | |
| ASH-S@90Model=VGG-162022.09 | 52.09 | 88.91 | — | |
| GAIA-ZMethods Space=Gradient, Backbone=Google BiT-S2023.11 | 52.92 | 86.89 | — | |
| KNN2022.12 | 53.97 | 85.01 | — | |
| ODINBackbone=MobileNet2021.11 | 54.2 | 85.81 | — | |
| ODINBackbone=MobileNet2023.08 | 54.2 | 85.81 | — | |
| KNNMethods Space=Feature, Backbone=Google BiT-S2023.11 | 54.32 | 84.59 | — | |
| Energy scoreModel=VGG-162022.09 | 54.33 | 88.17 | — | |
| KNNTraining scheme=Transfer learning, Model=ViT-B/16, ID accuracy=85.3%2023.09 | 54.45 | 87.62 | 96.93 | |
| GradNormMethods Space=Gradient, Backbone=Google BiT-S2023.11 | 54.7 | 86.31 | — | |
| ODINBackbone=ResNet-502021.11 | 56.48 | 85.41 | — | |
| ODINBackbone=ResNet-50, Pre-trained=ImageNet-1k2021.11 | 56.48 | 85.41 | — | |
| ODINBackbone=ResNet2023.08 | 56.48 | 85.41 | — | |
| ODIN2022.12 | 56.48 | 85.41 | — | |
| ASH-PBackbone=MobileNet2023.08 | 57.15 | 87.34 | — | |
| ReActMethods Space=Feature, Backbone=Google BiT-S2023.11 | 57.66 | 86.67 | — | |
| CaptionMethod Category=Textual OE, Level=Caption2023.10 | 58.35 | 85.53 | — | |
| EnergyBackbone=ResNet-502021.11 | 58.41 | 86.17 | — | |
| EnergyBackbone=ResNet-50, Pre-trained=ImageNet-1k2021.11 | 58.41 | 86.17 | — | |
| Energy scoreBackbone=ResNet2023.08 | 58.41 | 86.17 | — | |
| Energy2022.12 | 58.41 | 86.17 | — | |
| WordMethod Category=Textual OE, Level=Word2023.10 | 59.11 | 85.45 | — | |
| SSDTraining scheme=Transfer learning, Model=ViT-B/16, ID accuracy=85.3%2023.09 | 59.74 | 80.38 | 94.74 | |
| EnergyBackbone=MobileNet2021.11 | 62.39 | 84.91 | — | |
| Energy scoreBackbone=MobileNet2023.08 | 62.39 | 84.91 | — |