Out-of-Distribution Detection on ImageNet-1k ID iNaturalist OOD
0.23FPR95DAL-ASH
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| DAL-ASHTraining Data Setup=Using ID data and auxiliary OOD data, Backbone=ResNet-502023.11 | 0.23 | 99.85 | |
| POEMTraining Data Setup=Using ID data and auxiliary OOD data, Backbone=ResNet-502023.11 | 0.26 | 99.7 | |
| InterNegArchitecture=CLIP ViT-B/16, Modality=VLM-based, Training Requirement=no training on ID or extra data2026.03 | 0.4 | 99.79 | |
| DAL-EnergyTraining Data Setup=Using ID data and auxiliary OOD data, Backbone=ResNet-502023.11 | 0.47 | 99.82 | |
| ANTSTraining Protocol=Zero Shot (No Training Required), Encoder=ViT-B/162025.09 | 0.54 | 99.75 | |
| AdaNegArchitecture=CLIP ViT-B/16, Modality=VLM-based, Training Requirement=no training on ID or extra data2026.03 | 0.59 | 99.71 | |
| AdaNegTraining Protocol=Zero Shot (No Training Required), Encoder=ViT-B/162025.09 | 0.59 | 99.71 | |
| DOETraining Data Setup=Using ID data and auxiliary OOD data, Backbone=ResNet-502023.11 | 0.72 | 99.79 | |
| LAPTArchitecture=CLIP ViT-B/16, Modality=VLM-based, Training Requirement=requiring training on ID or extra data2026.03 | 1.16 | 99.63 | |
| LAPTTraining Protocol=Training-required (or with Fine-tuning), Encoder=ViT-B/162025.09 | 1.16 | 99.63 | |
| CSP+DCACArchitecture=CLIP-B/162026.01 | 1.2 | 99.62 | |
| CLIPScopeArchitecture=CLIP ViT-B/16, Modality=VLM-based, Training Requirement=no training on ID or extra data2026.03 | 1.29 | 99.61 | |
| CSPArchitecture=CLIP-B/162026.01 | 1.54 | 99.6 | |
| CSPArchitecture=CLIP ViT-B/16, Modality=VLM-based, Training Requirement=no training on ID or extra data2026.03 | 1.54 | 99.6 | |
| CSPTraining Protocol=Zero Shot (No Training Required), Encoder=ViT-B/162025.09 | 1.54 | 99.6 | |
| SynOODTraining Protocol=Training-required (or with Fine-tuning), Encoder=ViT-B/162025.09 | 1.57 | 99.57 | |
| CMATraining Protocol=Training-required (or with Fine-tuning), Encoder=ViT-B/162025.09 | 1.65 | 99.62 | |
| Energy-OETraining Data Setup=Using ID data and auxiliary OOD data, Backbone=ResNet-502023.11 | 1.88 | 99.49 | |
| NegLabelArchitecture=CLIP ViT-B/16, Modality=VLM-based, Training Requirement=no training on ID or extra data2026.03 | 1.91 | 99.49 | |
| NegLabelTraining Protocol=Zero Shot (No Training Required), Encoder=ViT-B/162025.09 | 1.91 | 99.49 | |
| OODDArchitecture=CLIP-B/162026.01 | 2.22 | 99.36 | |
| OODDTraining Protocol=Zero Shot (No Training Required), Encoder=ViT-B/162025.09 | 2.22 | 99.36 | |
| X-MahalanobisBackbone=ViT-B/16, Pre-training=ImageNet-21K, Evaluation Protocol=Fine-tuning2026.06 | 2.26 | 99.4 | |
| AdaNDArchitecture=CLIP-B/162026.01 | 4.19 | 98.91 | |
| Mahalanobis++Backbone=ViT-B/16, Pre-training=ImageNet-21K, Evaluation Protocol=Post-hoc2026.06 | 5.15 | 98.76 | |
| Relative Mahalanobis++Backbone=ViT-B/16, Pre-training=ImageNet-21K, Evaluation Protocol=Post-hoc2026.06 | 5.16 | 98.77 | |
| MM++Backbone=ViT-B/16, Pre-training=ImageNet-21K, Evaluation Protocol=Post-hoc, K=22026.06 | 5.71 | 98.43 | |
| DALTraining Data Setup=Using ID data and auxiliary OOD data, Backbone=ResNet-502023.11 | 5.83 | 99.09 | |
| NegPromptArchitecture=CLIP ViT-B/16, Modality=VLM-based, Training Requirement=requiring training on ID or extra data2026.03 | 6.32 | 98.73 | |
| ASH-S+DCACArchitecture=ResNet502026.01 | 6.69 | 98.54 | |
| ATOMTraining Data Setup=Using ID data and auxiliary OOD data, Backbone=ResNet-502023.11 | 7.07 | 98.25 | |
| OODDArchitecture=ResNet502026.01 | 7.13 | 98.71 | |
| OETraining Data Setup=Using ID data and auxiliary OOD data, Backbone=ResNet-502023.11 | 7.92 | 98.04 | |
| SUPREMEArchitecture=CLIP ViT-B/16, Modality=VLM-based, Training Requirement=requiring training on ID or extra data2026.03 | 8.27 | 98.29 | |
| Local-PromptArchitecture=CLIP ViT-B/16, Modality=VLM-based, Training Requirement=requiring training on ID or extra data2026.03 | 8.63 | 98.07 | |
| CMAArchitecture=CLIP-B/162026.01 | 8.88 | 98.19 | |
| ID-LikeArchitecture=CLIP ViT-B/16, Modality=VLM-based, Training Requirement=requiring training on ID or extra data2026.03 | 8.98 | 98.19 | |
| ID-LikeTraining Protocol=Training-required (or with Fine-tuning), Encoder=ViT-B/162025.09 | 8.98 | 98.19 | |
| Relative MahalanobisBackbone=ViT-B/16, Pre-training=ImageNet-21K, Evaluation Protocol=Post-hoc2026.06 | 10.62 | 97.61 | |
| GradOrthBackbone=ResNet2023.08 | 11.04 | 98 | |
| ASH-SBackbone=ResNet2023.08 | 11.49 | 97.87 | |
| ASH-SArchitecture=ResNet502026.01 | 11.49 | 97.87 | |
| DDCSArchitecture=ResNet502026.01 | 11.63 | 97.85 | |
| LINEArchitecture=ResNet502026.01 | 12.26 | 97.56 | |
| EOEArchitecture=CLIP ViT-B/16, Modality=VLM-based, Training Requirement=no training on ID or extra data2026.03 | 12.29 | 97.52 | |
| EOETraining Protocol=Zero Shot (No Training Required), Encoder=ViT-B/162025.09 | 12.29 | 97.52 | |
| MahalanobisBackbone=ViT-B/16, Pre-training=ImageNet-21K, Evaluation Protocol=Post-hoc2026.06 | 13.14 | 96.98 | |
| SCTArchitecture=CLIP-B/162026.01 | 13.94 | 95.86 | |
| SCTTraining Protocol=Training-required (or with Fine-tuning), Encoder=ViT-B/162025.09 | 13.94 | 95.86 | |
| ASH-BBackbone=ResNet2023.08 | 14.21 | 97.32 | |
| OSPCoOpArchitecture=CLIP-B/162026.01 | 15.25 | 97.13 | |
| ASH-S@90Model=DenseNet-1212022.09 | 15.53 | 97.03 | |
| VRA-PBackbone=ResNet2023.08 | 15.7 | 97.12 | |
| LoCoOpArchitecture=CLIP ViT-B/16, Modality=VLM-based, Training Requirement=requiring training on ID or extra data2026.03 | 16.05 | 96.86 | |
| CADRefArchitecture=ResNet502026.01 | 16.08 | 96.9 | |
| KNNBackbone=ViT-B/16, Pre-training=ImageNet-21K, Evaluation Protocol=Post-hoc2026.06 | 16.2 | 96.57 | |
| NPOSIn-distribution image requirement=Required2024.09 | 16.58 | 96.19 | |
| OptFSArchitecture=ResNet502026.01 | 16.79 | 96.88 | |
| VRA-DNBackbone=ResNet2023.08 | 16.82 | 96.92 | |
| ASH-B@90Model=DenseNet-1212022.09 | 18.22 | 96.36 | |
| DICE + ReActBackbone=ResNet2023.08 | 18.64 | 96.24 | |
| ReActBackbone=ViT-B/16, Pre-training=ImageNet-21K, Evaluation Protocol=Post-hoc2026.06 | 18.75 | 95.54 | |
| ReActArchitecture=ResNet502026.01 | 19.55 | 96.39 | |
| ReActBackbone=ResNet2023.08 | 20.38 | 96.22 | |
| LSNArchitecture=CLIP ViT-B/16, Modality=VLM-based, Training Requirement=requiring training on ID or extra data2026.03 | 21.56 | 95.83 | |
| LSNTraining Protocol=Training-required (or with Fine-tuning), Encoder=ViT-B/162025.09 | 21.56 | 95.83 | |
| CoVerArchitecture=CLIP ViT-B/16, Modality=VLM-based, Training Requirement=no training on ID or extra data2026.03 | 22.55 | 95.98 | |
| CoVerTraining Protocol=Zero Shot (No Training Required), Encoder=ViT-B/162025.09 | 22.55 | 95.98 | |
| CLIPNArchitecture=CLIP ViT-B/16, Modality=VLM-based, Training Requirement=requiring training on ID or extra data2026.03 | 23.94 | 95.27 | |
| CLIPNTraining Protocol=Training-required (or with Fine-tuning), Encoder=ViT-B/162025.09 | 23.94 | 95.27 | |
| DICEBackbone=ResNet2023.08 | 25.63 | 94.49 | |
| ASH-B@65Model=VGG-162022.09 | 25.98 | 94.2 | |
| EnergyArchitecture=CLIP ViT-B/16, Modality=Visual-based, Training Requirement=requiring training on ID or extra data2026.03 | 26.12 | 95.33 | |
| GradOrthBackbone=MobileNet2023.08 | 26.81 | 93.17 | |
| HFTTIn-distribution image requirement=Not required2024.09 | 27.44 | 93.27 | |
| VOSArchitecture=CLIP ViT-B/16, Modality=Visual-based, Training Requirement=requiring training on ID or extra data2026.03 | 28.99 | 94.62 | |
| KNNIn-distribution image requirement=Required2024.09 | 29.17 | 94.52 | |
| KNNArchitecture=CLIP ViT-B/16, Modality=Visual-based, Training Requirement=requiring training on ID or extra data2026.03 | 29.17 | 94.52 | |
| ASH-S@99Model=VGG-162022.09 | 29.3 | 93.18 | |
| LoCoOpTraining Protocol=Training-required (or with Fine-tuning), Encoder=ViT-B/162025.09 | 29.45 | 93.93 | |
| EnergyBackbone=ViT-B/16, Pre-training=ImageNet-21K, Evaluation Protocol=Post-hoc2026.06 | 29.73 | 91.59 | |
| ODINArchitecture=CLIP ViT-B/16, Modality=Visual-based, Training Requirement=requiring training on ID or extra data2026.03 | 30.22 | 94.65 | |
| ReActBackbone=ResNet-502026.02 | 30.9 | 94.6 | |
| ASH-BBackbone=MobileNet2023.08 | 31.46 | 94.28 | |
| MCMArchitecture=CLIP-B/162026.01 | 31.95 | 94.16 | |
| ViMArchitecture=CLIP ViT-B/16, Modality=Visual-based, Training Requirement=requiring training on ID or extra data2026.03 | 32.19 | 93.16 | |
| MCMArchitecture=CLIP ViT-B/16, Modality=VLM-based, Training Requirement=no training on ID or extra data2026.03 | 32.2 | 94.59 | |
| MCMTraining Protocol=Zero Shot (No Training Required), Encoder=ViT-B/162025.09 | 32.2 | 94.59 | |
| DICE + ReActBackbone=MobileNet2023.08 | 32.3 | 93.57 | |
| ASH-B@65Model=DenseNet-1212022.09 | 34.05 | 92.28 | |
| MCMIn-distribution image requirement=Not required2024.09 | 34.33 | 91.36 | |
| MSPBackbone=ViT-B/16, Pre-training=ImageNet-21K, Evaluation Protocol=Post-hoc2026.06 | 34.35 | 92.57 | |
| EnergyIn-distribution image requirement=Not required2024.09 | 34.7 | 90.55 | |
| MaxLogitIn-distribution image requirement=Not required2024.09 | 35.03 | 89.46 | |
| NegPromptTraining Protocol=Training-required (or with Fine-tuning), Encoder=ViT-B/162025.09 | 37.79 | 90.49 | |
| ASH-S@95Model=VGG-162022.09 | 38.36 | 92.98 | |
| ASH-SBackbone=MobileNet2023.08 | 39.1 | 91.94 | |
| Energy scoreModel=DenseNet-1212022.09 | 39.69 | 92.66 | |
| MSPIn-distribution image requirement=Required2024.09 | 40.17 | 89.76 | |
| FDBDBackbone=ResNet-502026.02 | 41 | 93.5 |