Out-of-Distribution Detection on iNaturalist
99.9AUROCNNGuide
Evaluation results
| Method | Links | |||||
|---|---|---|---|---|---|---|
| NNGuideBackbone=RegNet2025.01 | 99.9 | 1.8 | — | — | — | |
| MOODv2Backbone=BEiTv22025.01 | 99.6 | 1.8 | — | — | — | |
| MahalanobisModel=ViT-B/16, Source=feat+label2022.03 | 99.54 | 2.12 | — | — | — | |
| ViMModel=ViT-B/16, Source=feat+logit2022.03 | 99.41 | 2.6 | — | — | — | |
| ViMModel=Swin, Source=feat+logit2022.03 | 99.28 | 2.6 | — | — | — | |
| DisCoPatch-64Number of Patches=642025.01 | 99.1 | 3.6 | — | — | — | |
| ReActModel=ViT-B/16, Source=feat+logit2022.03 | 99 | 4.31 | — | — | — | |
| ResidualModel=Swin, Source=feat2022.03 | 98.89 | 4.81 | — | — | — | |
| DynProtoBackbone=ConvNeXt-B2026.04 | 98.75 | 4.34 | — | — | — | |
| DynProtoBackbone=ViT-B/162026.04 | 98.74 | 4.39 | — | — | — | |
| MahalanobisModel=Swin, Source=feat+label2022.03 | 98.69 | 5.43 | — | — | — | |
| EnergyModel=ViT-B/16, Source=logit2022.03 | 98.66 | 6.16 | — | — | — | |
| ODINModel=ViT-B/16, Source=prob+grad2022.03 | 98.57 | 6.58 | — | — | — | |
| MaxLogitModel=ViT-B/16, Source=logit2022.03 | 98.57 | 6.58 | — | — | — | |
| ResidualModel=ViT-B/16, Source=feat2022.03 | 98.57 | 6.63 | — | — | — | |
| Catalyst(μ) + ReActBackbone=ResNet-502026.02 | 98.19 | 8.54 | — | — | — | |
| MOSgrouping=taxonomy-based2021.05 | 98.15 | 9.28 | 99.62 | — | — | |
| MOSBackbone=ResNetv2-101, Pre-trained=ImageNet2023.02 | 98.15 | 9.28 | — | — | — | |
| Catalyst(σ) + ReActBackbone=ResNet-502026.02 | 98.06 | 9.1 | — | — | — | |
| SCALEBackbone=ResNet-502026.02 | 98.02 | 10.37 | — | — | — | |
| SCALEBackbone=ResNet-502025.01 | 98 | 9.5 | — | — | — | |
| GradOrthBackbone=ResNet-502026.02 | 98 | 11.04 | — | — | — | |
| Catalyst(m) + ReActBackbone=ResNet-502026.02 | 97.97 | 9.71 | — | — | — | |
| ASHBackbone=ResNet-502025.01 | 97.9 | 11.5 | — | — | — | |
| ASHBackbone=ResNet-502026.02 | 97.87 | 11.52 | — | — | — | |
| VRA+Backbone=ResNetv2-101, Pre-trained=ImageNet2023.02 | 97.7 | 20.81 | — | — | — | |
| BATSBackbone=ResNet-502026.02 | 97.67 | 12.57 | — | — | — | |
| FDBDBackbone=ResNet-502025.01 | 97.6 | 12.4 | — | — | — | |
| ReActModel=Swin, Source=feat2022.03 | 97.51 | 9.98 | — | — | — | |
| LAPSBackbone=ResNet-502026.02 | 97.5 | 12.72 | — | — | — | |
| WeiPer+KLDBackbone=ResNet50, Training Loss=Cross Entropy, Preprocessing=Standard Preprocessing2024.05 | 97.49 | 13.59 | — | — | — | |
| NN-GuideBackbone=ResNet-502026.02 | 97.47 | 12.02 | — | — | — | |
| VIMBackbone=ViT-B/162026.04 | 97.18 | 12.55 | — | — | — | |
| VRABackbone=ResNet-50, Pre-trained=ImageNet2023.02 | 97.12 | 15.7 | — | — | — | |
| VRA+Backbone=ResNet-50, Pre-trained=ImageNet2023.02 | 97.08 | 15.48 | — | — | — | |
| ASHBackbone=ResNet50, Training Loss=Cross Entropy, Preprocessing=Standard Preprocessing2024.05 | 97.07 | 14.04 | — | — | — | |
| ReAct+DICEBackbone=ResNet-502026.02 | 97.06 | 14.9 | — | — | — | |
| NNGuideBackbone=ResNet-502025.01 | 96.9 | 14.3 | — | — | — | |
| Adaptive Multi-prompt Contrastive NetworkShot=82025.06 | 96.89 | 18.17 | — | — | — | |
| KL MatchingModel=ViT-B/16, Source=prob2022.03 | 96.88 | 14.79 | — | — | — | |
| LAPSBackbone=MobileNet-v22026.02 | 96.76 | 18.82 | — | — | — | |
| GL-MCMShot=02025.06 | 96.71 | 15.16 | — | — | — | |
| NACBackbone=ResNet50, Training Loss=Cross Entropy, Preprocessing=Standard Preprocessing2024.05 | 96.52 | — | — | — | — | |
| Catalyst(μ)Backbone=ResNet-502026.02 | 96.46 | 18.02 | — | — | — | |
| ReActBackbone=ResNet-502026.02 | 96.37 | 19.73 | — | — | — | |
| ReActBackbone=ResNet50, Training Loss=Cross Entropy, Preprocessing=Standard Preprocessing2024.05 | 96.34 | 16.72 | — | — | — | |
| ReActBackbone=ResNet-50, Pre-trained=ImageNet2023.02 | 96.22 | 20.38 | — | — | — | |
| Catalyst(σ)Backbone=ResNet-502026.02 | 96.21 | 19.05 | — | — | — | |
| NPOSEvaluation Protocol=fine-tuned2023.03 | 96.19 | 16.58 | — | — | — | |
| NPOSShot=Full2025.06 | 96.19 | 16.58 | — | — | — | |
| Catalyst(m)Backbone=ResNet-502026.02 | 96.18 | 19 | — | — | — | |
| Adaptive Multi-prompt Contrastive NetworkShot=12025.06 | 96.18 | 18.84 | — | — | — | |
| MSPModel=ViT-B/16, Source=prob2022.03 | 96.11 | 19.04 | — | — | — | |
| RMDSBackbone=ViT16-B, Training Objective=Cross Entropy, Preprocessing=standard2024.05 | 96.1 | 19.47 | — | — | — | |
| MDSBackbone=ViT16-B, Training Objective=Cross Entropy, Preprocessing=standard2024.05 | 96.01 | 20.64 | — | — | — | |
| RankFeatBackbone=ResNetv2-1012025.01 | 96 | 13 | — | — | — | |
| SCTShot=82025.06 | 95.82 | 18.65 | — | — | — | |
| WeiPer+ReActBackbone=ResNet50, Training Loss=Cross Entropy, Preprocessing=Standard Preprocessing2024.05 | 95.75 | 21.03 | — | — | — | |
| MaxLogitModel=Swin, Source=logit2022.03 | 95.72 | 15.41 | — | — | — | |
| VIMBackbone=ViT16-B, Training Objective=Cross Entropy, Preprocessing=standard2024.05 | 95.72 | 17.59 | — | — | — | |
| SCTShot=12025.06 | 95.7 | 19.16 | — | — | — | |
| VRABackbone=ResNetv2-101, Pre-trained=ImageNet2023.02 | 95.68 | 27.26 | — | — | — | |
| Catalyst(m) + ReActBackbone=MobileNet-v22026.02 | 95.66 | 24.08 | — | — | — | |
| EnergyModel=Swin, Source=logit2022.03 | 95.22 | 15.47 | — | — | — | |
| Catalyst(σ) + ReActBackbone=MobileNet-v22026.02 | 95.12 | 27.21 | — | — | — | |
| OpenMaxBackbone=ViT16-B, Training Objective=Cross Entropy, Preprocessing=standard2024.05 | 94.93 | 19.62 | — | — | — | |
| Catalyst(μ) + ReActBackbone=MobileNet-v22026.02 | 94.89 | 28.19 | — | — | — | |
| LoCoOpShot=82025.06 | 94.86 | 27.45 | — | — | — | |
| KL MatchingModel=Swin, Source=prob2022.03 | 94.77 | 27.62 | — | — | — | |
| MSPModel=Swin, Source=prob2022.03 | 94.76 | 23.19 | — | — | — | |
| Catalyst(m)Backbone=MobileNet-v22026.02 | 94.7 | 28.78 | — | — | — | |
| EnergyEvaluation Protocol=fine-tuned2023.03 | 94.68 | 29.75 | — | — | — | |
| SeTARShot=02025.06 | 94.67 | 26.92 | — | — | — | |
| ODINEvaluation Protocol=fine-tuned2023.03 | 94.65 | 30.22 | — | — | — | |
| ODINShot=Full2025.06 | 94.65 | 30.22 | — | — | — | |
| Catalyst(σ)Backbone=MobileNet-v22026.02 | 94.63 | 29.25 | — | — | — | |
| VOS+Evaluation Protocol=fine-tuned2023.03 | 94.62 | 28.99 | — | — | — | |
| VOSEvaluation Protocol=fine-tuned2023.03 | 94.53 | 31.65 | — | — | — | |
| DICEBackbone=ResNet-502026.02 | 94.53 | 26.48 | — | — | — | |
| KNNEvaluation Protocol=fine-tuned2023.03 | 94.52 | 29.17 | — | — | — | |
| KNNShot=Full2025.06 | 94.52 | 29.17 | — | — | — | |
| CADRefBackbone=ConvNeXt-B2026.04 | 94.5 | 33.08 | — | — | — | |
| DICEBackbone=ResNet-50, Pre-trained=ImageNet2023.02 | 94.49 | 25.63 | — | — | — | |
| SCALEBackbone=MobileNet-v22026.02 | 94.46 | 30.09 | — | — | — | |
| MCMEvaluation Protocol=zero-shot2023.03 | 94.41 | 32.08 | — | — | — | |
| BATSBackbone=MobileNet-v22026.02 | 94.33 | 31.56 | — | — | — | |
| ODINModel=Swin, Source=prob+grad2022.03 | 94.24 | 19.65 | — | — | — | |
| MCMShot=02025.06 | 94.17 | 31.86 | — | — | — | |
| LoCoOpShot=12025.06 | 94.05 | 28.81 | — | — | — | |
| GradNormBackbone=ResNet-502026.02 | 93.97 | 23.73 | — | — | — | |
| GradNormBackbone=ResNet50, Training Loss=Cross Entropy, Preprocessing=Standard Preprocessing2024.05 | 93.89 | 32.03 | — | — | — | |
| Catalyst(μ)Backbone=MobileNet-v22026.02 | 93.84 | 33.47 | — | — | — | |
| CADRefBackbone=ViT-B/162026.04 | 93.81 | 38.45 | — | — | — | |
| ReAct+DICEBackbone=MobileNet-v22026.02 | 93.76 | 31.68 | — | — | — | |
| NACBackbone=ViT16-B, Training Objective=Cross Entropy, Preprocessing=standard2024.05 | 93.72 | — | — | — | — | |
| fDBDBackbone=ResNet-502026.02 | 93.67 | 40.24 | — | — | — | |
| VIMBackbone=ConvNeXt-B2026.04 | 93.63 | 40.84 | — | — | — | |
| SHEBackbone=ViT16-B, Training Objective=Cross Entropy, Preprocessing=standard2024.05 | 93.57 | 22.16 | — | — | — | |
| GENBackbone=ViT16-B, Training Objective=Cross Entropy, Preprocessing=standard2024.05 | 93.54 | 22.92 | — | — | — | |
| GradOrthBackbone=MobileNet-v22026.02 | 93.17 | 26.81 | — | — | — |