Out-of-Distribution Detection on OpenImage-O
97.7AUROCNNGuide
Evaluation results
| Method | Links | ||
|---|---|---|---|
| NNGuideBackbone=RegNet2025.01 | 97.7 | 10.8 | |
| ViMModel=ViT-B/16, Source=feat+logit2022.03 | 97.61 | 12.61 | |
| MahalanobisModel=ViT-B/16, Source=feat+label2022.03 | 97.48 | 13.54 | |
| MOODv2Backbone=BEiTv22025.01 | 97.4 | 13.6 | |
| ReActModel=ViT-B/16, Source=feat+logit2022.03 | 97.38 | 13.5 | |
| EnergyModel=ViT-B/16, Source=logit2022.03 | 97.11 | 14.04 | |
| MaxLogitModel=ViT-B/16, Source=logit2022.03 | 96.87 | 15.68 | |
| ODINModel=ViT-B/16, Source=prob+grad2022.03 | 96.86 | 15.68 | |
| ViMModel=Swin, Source=feat+logit2022.03 | 96.04 | 23.88 | |
| ResidualModel=Swin, Source=feat2022.03 | 94.64 | 32.19 | |
| MahalanobisModel=Swin, Source=feat+label2022.03 | 94.57 | 33.41 | |
| DisCoPatch-64Number of Patches=642025.01 | 94.4 | 29.7 | |
| SCALEBackbone=ResNet-502025.01 | 94 | 28.2 | |
| KL MatchingModel=ViT-B/16, Source=prob2022.03 | 93.8 | 28.49 | |
| ReActModel=Swin, Source=feat2022.03 | 93.58 | 23.07 | |
| MM++Backbone=ConvNeXt-T, In-Distribution Dataset=ImageNet-LT2026.06 | 93.4 | 37.69 | |
| ASHBackbone=ResNet50, Training Loss=Cross Entropy, Preprocessing=Standard Preprocessing2024.05 | 93.26 | 29.15 | |
| rMaha++Backbone=ConvNeXt-T, In-Distribution Dataset=ImageNet-LT2026.06 | 93.14 | 39.47 | |
| Maha++Backbone=ConvNeXt-T, In-Distribution Dataset=ImageNet-LT2026.06 | 93.12 | 39.51 | |
| WeiPer+KLDBackbone=ResNet50, Training Loss=Cross Entropy, Preprocessing=Standard Preprocessing2024.05 | 92.94 | 30.49 | |
| ASHBackbone=ResNet-502025.01 | 92.8 | 32.7 | |
| ResidualModel=ViT-B/16, Source=feat2022.03 | 92.72 | 32.63 | |
| rMahaBackbone=ConvNeXt-T, In-Distribution Dataset=ImageNet-LT2026.06 | 92.61 | 42.71 | |
| MSPModel=ViT-B/16, Source=prob2022.03 | 92.53 | 34.18 | |
| NNGuideBackbone=ResNet-502025.01 | 92.4 | 35.4 | |
| RankFeatBackbone=ResNetv2-1012025.01 | 92.4 | 33.3 | |
| MDSBackbone=ViT16-B, Training Objective=Cross Entropy, Preprocessing=standard2024.05 | 92.38 | 30.35 | |
| RMDSBackbone=ViT16-B, Training Objective=Cross Entropy, Preprocessing=standard2024.05 | 92.32 | 29.57 | |
| VIMBackbone=ViT16-B, Training Objective=Cross Entropy, Preprocessing=standard2024.05 | 92.18 | 29.61 | |
| KL MatchingModel=Swin, Source=prob2022.03 | 91.92 | 40.05 | |
| MaxLogitModel=Swin, Source=logit2022.03 | 91.91 | 26.79 | |
| ReActBackbone=ResNet50, Training Loss=Cross Entropy, Preprocessing=Standard Preprocessing2024.05 | 91.87 | 32.58 | |
| FDBDBackbone=ResNet-502025.01 | 91.7 | 35.6 | |
| WeiPer+ReActBackbone=ResNet50, Training Loss=Cross Entropy, Preprocessing=Standard Preprocessing2024.05 | 91.64 | 33.53 | |
| NACBackbone=ViT16-B, Training Objective=Cross Entropy, Preprocessing=standard2024.05 | 91.58 | — | |
| MahaBackbone=ConvNeXt-T, In-Distribution Dataset=ImageNet-LT2026.06 | 91.57 | 48.96 | |
| ViMModel=BiT-S-R101x1, Source=feat+logit2022.03 | 91.54 | 43.96 | |
| NACBackbone=ResNet50, Training Loss=Cross Entropy, Preprocessing=Standard Preprocessing2024.05 | 91.45 | — | |
| ODINModel=Swin, Source=prob+grad2022.03 | 91.38 | 28.42 | |
| MSPModel=Swin, Source=prob2022.03 | 91.35 | 34.96 | |
| SHEBackbone=ViT16-B, Training Objective=Cross Entropy, Preprocessing=standard2024.05 | 91.04 | 33.57 | |
| EnergyModel=Swin, Source=logit2022.03 | 90.93 | 27.58 | |
| ViMModel=Res50d, Source=feat+logit2022.03 | 90.76 | 50.45 | |
| VIMBackbone=ResNet50, Training Loss=Cross Entropy, Preprocessing=Standard Preprocessing2024.05 | 90.5 | 32.82 | |
| GENBackbone=ViT16-B, Training Objective=Cross Entropy, Preprocessing=standard2024.05 | 90.27 | 35.47 | |
| KNNBackbone=ViT16-B, Training Objective=Cross Entropy, Preprocessing=standard2024.05 | 89.86 | 34.82 | |
| MahalanobisModel=Res50d, Source=feat+label2022.03 | 89.52 | 55.91 | |
| WeiPer+KLDBackbone=ViT16-B, Training Objective=Cross Entropy, Preprocessing=standard2024.05 | 89.51 | 39.05 | |
| ViMModel=RepVGG, Source=feat+logit2022.03 | 89.27 | 52.4 | |
| GENBackbone=ResNet50, Training Loss=Cross Entropy, Preprocessing=Standard Preprocessing2024.05 | 89.26 | 34.5 | |
| MLSBackbone=ResNet50, Training Loss=Cross Entropy, Preprocessing=Standard Preprocessing2024.05 | 89.17 | 37.88 | |
| ViMModel=DeiT, Source=feat+logit2022.03 | 89.13 | 64.58 | |
| EBOBackbone=ResNet50, Training Loss=Cross Entropy, Preprocessing=Standard Preprocessing2024.05 | 89.06 | 38.09 | |
| MahalanobisModel=DeiT, Source=feat+label2022.03 | 89.03 | 66.51 | |
| KL MatchingModel=BiT-S-R101x1, Source=prob2022.03 | 88.96 | 51.51 | |
| ReActModel=BiT-S-R101x1, Source=feat+logit2022.03 | 88.94 | 54.97 | |
| WeiPer+MSPBackbone=ResNet50, Training Loss=Cross Entropy, Preprocessing=Standard Preprocessing2024.05 | 88.94 | 39.75 | |
| WeiPer+ReActBackbone=ViT16-B, Training Objective=Cross Entropy, Preprocessing=standard2024.05 | 88.56 | 43.26 | |
| DICEBackbone=ResNet50, Training Loss=Cross Entropy, Preprocessing=Standard Preprocessing2024.05 | 88.26 | 47.83 | |
| ODINBackbone=ResNet50, Training Loss=Cross Entropy, Preprocessing=Standard Preprocessing2024.05 | 88.23 | 46.67 | |
| WeiPer+MSPBackbone=ViT16-B, Training Objective=Cross Entropy, Preprocessing=standard2024.05 | 88.15 | 46.3 | |
| ResidualModel=DeiT, Source=feat2022.03 | 88.07 | 69.21 | |
| ResidualModel=Res50d, Source=feat2022.03 | 87.64 | 59.65 | |
| OpenMaxBackbone=ResNet50, Training Loss=Cross Entropy, Preprocessing=Standard Preprocessing2024.05 | 87.62 | 37.39 | |
| KL MatchingModel=DeiT, Source=prob2022.03 | 87.49 | 60.66 | |
| OpenMaxBackbone=ViT16-B, Training Objective=Cross Entropy, Preprocessing=standard2024.05 | 87.36 | 73.82 | |
| KL MatchingModel=Res50d, Source=prob2022.03 | 87.31 | 60.58 | |
| KLMBackbone=ResNet50, Training Loss=Cross Entropy, Preprocessing=Standard Preprocessing2024.05 | 87.3 | 48.89 | |
| TempScaleBackbone=ResNet50, Training Loss=Cross Entropy, Preprocessing=Standard Preprocessing2024.05 | 87.22 | 45.4 | |
| KNNBackbone=ResNet50, Training Loss=Cross Entropy, Preprocessing=Standard Preprocessing2024.05 | 87.04 | 44.27 | |
| KLMBackbone=ViT16-B, Training Objective=Cross Entropy, Preprocessing=standard2024.05 | 87.03 | 51.75 | |
| KL MatchingModel=RepVGG, Source=prob2022.03 | 86.8 | 57.48 | |
| SHEBackbone=ResNet50, Training Loss=Cross Entropy, Preprocessing=Standard Preprocessing2024.05 | 86.52 | 55.02 | |
| RMDSBackbone=ResNet50, Training Loss=Cross Entropy, Preprocessing=Standard Preprocessing2024.05 | 85.84 | 40.27 | |
| MahalanobisModel=RepVGG, Source=feat+label2022.03 | 85.71 | 64.93 | |
| MaxLogitModel=BiT-S-R101x1, Source=logit2022.03 | 85.67 | 72.68 | |
| ODINModel=BiT-S-R101x1, Source=prob+grad2022.03 | 85.64 | 72.83 | |
| ReActModel=Res50d, Source=feat2022.03 | 85.3 | 60.79 | |
| ODINModel=RepVGG, Source=prob+grad2022.03 | 85.22 | 63.48 | |
| MSPModel=RepVGG, Source=prob2022.03 | 85.06 | 63.36 | |
| TempScaleBackbone=ViT16-B, Training Objective=Cross Entropy, Preprocessing=standard2024.05 | 85.04 | 59.98 | |
| MSPBackbone=ResNet50, Training Loss=Cross Entropy, Preprocessing=Standard Preprocessing2024.05 | 84.86 | 50.13 | |
| MSPBackbone=ViT16-B, Training Objective=Cross Entropy, Preprocessing=standard2024.05 | 84.86 | 56.19 | |
| GradNormBackbone=ResNet50, Training Loss=Cross Entropy, Preprocessing=Standard Preprocessing2024.05 | 84.82 | 68.46 | |
| MaxLogitModel=RepVGG, Source=logit2022.03 | 84.81 | 65.04 | |
| EnergyModel=BiT-S-R101x1, Source=logit2022.03 | 84.77 | 73.42 | |
| MSPModel=Res50d, Source=prob2022.03 | 84.5 | 63.53 | |
| ReActBackbone=ViT16-B, Training Objective=Cross Entropy, Preprocessing=standard2024.05 | 84.29 | 57.67 | |
| MSPModel=BiT-S-R101x1, Source=prob2022.03 | 84.16 | 73.72 | |
| MSPModel=DeiT, Source=prob2022.03 | 84.04 | 62.03 | |
| EnergyModel=RepVGG, Source=logit2022.03 | 83.64 | 69.92 | |
| MahalanobisModel=BiT-S-R101x1, Source=feat+label2022.03 | 83.1 | 64.32 | |
| NNGuideBackbone=MobileNetV22025.01 | 82.9 | 63.3 | |
| MSPBackbone=ConvNeXt-T, In-Distribution Dataset=ImageNet-LT2026.06 | 82.83 | 61.63 | |
| ResidualModel=RepVGG, Source=feat2022.03 | 82.51 | 65.13 | |
| DICEBackbone=ViT16-B, Training Objective=Cross Entropy, Preprocessing=standard2024.05 | 82.22 | 52.57 | |
| MLSBackbone=ViT16-B, Training Objective=Cross Entropy, Preprocessing=standard2024.05 | 81.6 | 85.82 | |
| ODINModel=Res50d, Source=prob+grad2022.03 | 81.53 | 64.49 | |
| MaxLogitModel=Res50d, Source=logit2022.03 | 81.5 | 65.5 | |
| ResidualModel=BiT-S-R101x1, Source=feat2022.03 | 80.58 | 67.85 |