OOD Detection on iNaturalist (OOD) vs ImageNet-1k (ID) (test)
0.2FPR95DynProto
Evaluation results
| Method | Links | ||
|---|---|---|---|
| DynProtoBackbone=CLIP-B/162026.04 | 0.2 | 99.96 | |
| TANLEncoder=VITB/16 CLIP, Protocol=Training-free & test-time adaptation methods2026.03 | 0.42 | 99.84 | |
| TTLExtra Resources=No, Category=Test-time adaptation2026.04 | 0.42 | 99.87 | |
| CoEvoCSPEncoder=CLIP ViT-B/16, Method Category=Training-free2026.01 | 0.46 | 99.82 | |
| CSPBackbone=CLIP-B/162026.04 | 0.46 | 99.9 | |
| CoEvoNegLabelEncoder=CLIP ViT-B/16, Method Category=Training-free2026.01 | 0.53 | 99.81 | |
| NegLabelBackbone=CLIP-B/162026.04 | 0.57 | 99.86 | |
| AdaNegEncoder=CLIP ViT-B/16, Method Category=Training-free2026.01 | 0.59 | 99.71 | |
| AdaNegEncoder=VITB/16 CLIP, Protocol=Training-free & test-time adaptation methods2026.03 | 0.59 | 99.71 | |
| OODDEncoder=VITB/16 CLIP, Protocol=Training-free & test-time adaptation methods2026.03 | 0.85 | 99.79 | |
| AdaNegExtra Resources=Yes, Category=Test-time adaptation2026.04 | 0.9 | 99.69 | |
| DualCnstEncoder=VITB/16 CLIP, Protocol=Training-free & non-adaptive methods2026.03 | 0.98 | 99.7 | |
| Peng et al.Encoder=VITB/16 CLIP, Protocol=Training-free & non-adaptive methods2026.03 | 1.04 | 99.64 | |
| LAPTEncoder=CLIP ViT-B/16, Method Category=Training-based2026.01 | 1.16 | 99.63 | |
| LAPTEncoder=VITB/16 CLIP, Protocol=Methods requiring training (or fine-tuning)2026.03 | 1.16 | 99.63 | |
| CLIPScopeEncoder=VITB/16 CLIP, Protocol=Training-free & test-time adaptation methods2026.03 | 1.29 | 99.61 | |
| CSPEncoder=CLIP ViT-B/16, Method Category=Training-free2026.01 | 1.54 | 99.6 | |
| CSPEncoder=VITB/16 CLIP, Protocol=Training-free & non-adaptive methods2026.03 | 1.54 | 99.6 | |
| CSPExtra Resources=Yes, Category=Post-hoc2026.04 | 1.54 | 99.6 | |
| SynOODEncoder=VITB/16 CLIP, Protocol=Methods requiring training (or fine-tuning)2026.03 | 1.57 | 99.57 | |
| NegRefineExtra Resources=Yes, Category=Post-hoc2026.04 | 1.61 | 99.53 | |
| CMA [30]Encoder=VITB/16 CLIP, Protocol=Methods requiring training (or fine-tuning)2026.03 | 1.65 | 99.62 | |
| NegLabelEncoder=CLIP ViT-B/16, Method Category=Training-free2026.01 | 1.91 | 99.49 | |
| AdaNDEncoder=VITB/16 CLIP, Protocol=Methods requiring training (or fine-tuning)2026.03 | 1.91 | 99.49 | |
| NegLabelEncoder=VITB/16 CLIP, Protocol=Training-free & non-adaptive methods2026.03 | 1.91 | 99.49 | |
| NeglabelExtra Resources=Yes, Category=Post-hoc2026.04 | 2 | 99.47 | |
| OODDExtra Resources=No, Category=Test-time adaptation2026.04 | 2.13 | 99.39 | |
| LoCoOpBackbone=CLIP-B/162026.04 | 3.81 | 99.04 | |
| AdaNDBackbone=CLIP-B/162026.04 | 4.68 | 98.59 | |
| TagOODEncoder=VITB/16 CLIP, Protocol=Methods requiring training (or fine-tuning)2026.03 | 5 | 98.97 | |
| MoFEExtra Resources=No, Category=Training-based2026.04 | 5.19 | 97.28 | |
| NegPromptEncoder=CLIP ViT-B/16, Method Category=Training-based2026.01 | 6.32 | 98.73 | |
| NegPromptEncoder=VITB/16 CLIP, Protocol=Methods requiring training (or fine-tuning)2026.03 | 6.32 | 98.73 | |
| CoOpBackbone=CLIP-B/162026.04 | 7 | 98.45 | |
| AdaNDExtra Resources=No, Category=Test-time adaptation2026.04 | 7.06 | 98.5 | |
| FABackbone=CLIP-B/162026.04 | 7.26 | 98.44 | |
| Local-PromptExtra Resources=No, Category=Training-based2026.04 | 8.62 | 98.06 | |
| MOSBackbone=BiT, Training Protocol=Fine-tuned2022.11 | 9.28 | 98.15 | |
| APTExtra Resources=Yes, Category=Training-based2026.04 | 9.7 | 97.79 | |
| EnergyBackbone=CLIP-L, Training Protocol=Fine-tuned2022.11 | 10.62 | 97.52 | |
| LINePost-hoc=true2023.03 | 12.26 | 97.56 | |
| EOEEncoder=VITB/16 CLIP, Protocol=Training-free & non-adaptive methods2026.03 | 12.29 | 97.52 | |
| MCMBackbone=CLIP-B/162026.04 | 12.87 | 97.3 | |
| FAExtra Resources=No, Category=Training-based2026.04 | 13.37 | 96.8 | |
| FAGLEncoder=VITB/16 CLIP, Protocol=Methods requiring training (or fine-tuning)2026.03 | 14.49 | 96.48 | |
| Fort et al.Backbone=ViT-B, Training Protocol=Fine-tuned2022.11 | 15.07 | 96.64 | |
| GL-MCMExtra Resources=No, Category=Post-hoc2026.04 | 15.09 | 96.72 | |
| OSPCoOpExtra Resources=Yes, Category=Training-based2026.04 | 15.25 | 97.13 | |
| Fort et al.Backbone=ViT-L, Training Protocol=Fine-tuned2022.11 | 15.74 | 96.51 | |
| LoCoOpEncoder=CLIP ViT-B/16, Method Category=Training-based2026.01 | 16.05 | 96.86 | |
| LoCoOpEncoder=VITB/16 CLIP, Protocol=Methods requiring training (or fine-tuning)2026.03 | 16.05 | 96.86 | |
| LoCoOpExtra Resources=No, Category=Training-based2026.04 | 16.05 | 96.86 | |
| NPOSEncoder=CLIP ViT-B/16, Method Category=Training-based2026.01 | 16.58 | 96.19 | |
| NPOSEncoder=VITB/16 CLIP, Protocol=Methods requiring training (or fine-tuning)2026.03 | 16.58 | 96.19 | |
| DICE + ReActPost-hoc=true2023.03 | 18.64 | 96.24 | |
| CLIPNExtra Resources=Yes, Category=Post-hoc2026.04 | 19.17 | 96.17 | |
| IDLikeExtra Resources=No, Category=Training-based2026.04 | 19.23 | 96.7 | |
| ReActPost-hoc=true2023.03 | 20.38 | 96.22 | |
| LSNEncoder=VITB/16 CLIP, Protocol=Methods requiring training (or fine-tuning)2026.03 | 21.56 | 95.83 | |
| EnergyBackbone=CLIP-B, Training Protocol=Fine-tuned2022.11 | 21.59 | 95.99 | |
| CoVerEncoder=VITB/16 CLIP, Protocol=Training-free & non-adaptive methods2026.03 | 22.55 | 95.98 | |
| CMA [37]Encoder=VITB/16 CLIP, Protocol=Training-free & non-adaptive methods2026.03 | 23.84 | 96.89 | |
| CMAExtra Resources=Yes, Category=Post-hoc2026.04 | 23.84 | 96.89 | |
| CLIPNEncoder=CLIP ViT-B/16, Method Category=Training-based2026.01 | 23.94 | 95.27 | |
| CLIPNEncoder=VITB/16 CLIP, Protocol=Methods requiring training (or fine-tuning)2026.03 | 23.94 | 95.27 | |
| DICEPost-hoc=true2023.03 | 25.63 | 94.49 | |
| EnergyEncoder=CLIP ViT-B/16, Method Category=Training-based2026.01 | 26.12 | 95.33 | |
| EnergyEncoder=VITB/16 CLIP, Protocol=Methods requiring training (or fine-tuning)2026.03 | 26.12 | 95.33 | |
| MCMBackbone=CLIP-L, Training Protocol=Zero-shot2022.11 | 28.38 | 94.95 | |
| VOSEncoder=VITB/16 CLIP, Protocol=Methods requiring training (or fine-tuning)2026.03 | 28.99 | 94.62 | |
| KNNEncoder=VITB/16 CLIP, Protocol=Methods requiring training (or fine-tuning)2026.03 | 29.17 | 94.52 | |
| ODINEncoder=VITB/16 CLIP, Protocol=Methods requiring training (or fine-tuning)2026.03 | 30.22 | 94.65 | |
| MCMBackbone=CLIP-B, Training Protocol=Zero-shot2022.11 | 30.91 | 94.61 | |
| MCMExtra Resources=No, Category=Post-hoc2026.04 | 30.92 | 94.61 | |
| ViMEncoder=VITB/16 CLIP, Protocol=Methods requiring training (or fine-tuning)2026.03 | 32.19 | 93.16 | |
| MCMEncoder=CLIP ViT-B/16, Method Category=Training-free2026.01 | 32.2 | 94.59 | |
| MCMEncoder=VITB/16 CLIP, Protocol=Training-free & non-adaptive methods2026.03 | 32.2 | 94.59 | |
| MSPBackbone=CLIP-L, Training Protocol=Fine-tuned2022.11 | 34.54 | 92.62 | |
| MSPBackbone=CLIP-B, Training Protocol=Fine-tuned2022.11 | 40.89 | 88.63 | |
| ODINPost-hoc=true2023.03 | 47.66 | 89.66 | |
| MSPPost-hoc=true2023.03 | 54.99 | 87.74 | |
| EnergyPost-hoc=true2023.03 | 55.72 | 89.95 | |
| MSPEncoder=CLIP ViT-B/16, Method Category=Training-based2026.01 | 58.36 | 87.44 | |
| MSPEncoder=VITB/16 CLIP, Protocol=Methods requiring training (or fine-tuning)2026.03 | 58.36 | 87.44 | |
| EnergyEncoder=CLIP ViT-B/16, Method Category=Training-free2026.01 | 81.08 | 85.09 | |
| GradNormEncoder=CLIP ViT-B/16, Method Category=Training-based2026.01 | 81.5 | 72.56 | |
| GradNormEncoder=VITB/16 CLIP, Protocol=Methods requiring training (or fine-tuning)2026.03 | 81.5 | 72.56 | |
| ZOCEncoder=CLIP ViT-B/16, Method Category=Training-based2026.01 | 87.3 | 86.09 | |
| ZOCEncoder=VITB/16 CLIP, Protocol=Methods requiring training (or fine-tuning)2026.03 | 87.3 | 86.09 | |
| MahalanobisPost-hoc=true2023.03 | 97 | 52.65 |