Out-of-Distribution Detection on CIFAR-100 (ID) vs CIFAR-10 (OOD) (test)
96.3AUROCGeneralized (Neg,Vol, InfoNCE)
Evaluation results
| Method | Links | |||
|---|---|---|---|---|
| Generalized (Neg,Vol, InfoNCE)Backbone=RepVGG-A22026.03 | 96.3 | — | 26 | |
| Single (Vol)Backbone=RepVGG-A22026.03 | 96.2 | — | 22.2 | |
| Single (Neg,Vol)Backbone=RepVGG-A22026.03 | 95.7 | — | 28.9 | |
| Generalized (Neg,Vol)Backbone=RepVGG-A22026.03 | 95.4 | — | 34.5 | |
| Single (Neg,Vol, InfoNCE)Backbone=RepVGG-A22026.03 | 95.2 | — | 31.3 | |
| Generalized (Vol)Backbone=RepVGG-A22026.03 | 95.1 | — | 35.5 | |
| k-NNBackbone=DINOv2, k-value=502026.05 | 93 | — | 35.2 | |
| E2 normed-forkBackbone=DINOv2, Embedding Dimension=768-d2026.05 | 92.3 | — | 33.3 | |
| Single (Neg)Backbone=RepVGG-A22026.03 | 91.8 | — | 49.9 | |
| EncMin2L2026.05 | 89.2 | — | 42.8 | |
| Generalized (Neg)Backbone=RepVGG-A22026.03 | 88.2 | — | 63.8 | |
| LoRA-EnsembleEnsemble size=16, LoRA=true2024.05 | 82.1 | 80.4 | 54.1 | |
| Single Network with LoRAEnsemble size=1, LoRA=true2024.05 | 80 | 78 | 57.3 | |
| Energy (T=1)Backbone=RepVGG-A22026.03 | 79.9 | — | 56.4 | |
| MSPBackbone=RepVGG-A22026.03 | 79.8 | — | 56.4 | |
| Split-EnsembleEnsemble size=162024.05 | 79.2 | 81.7 | 78.5 | |
| DM Dual Threshold2026.05 | 79.1 | — | — | |
| Explicit EnsembleEnsemble size=162024.05 | 79 | 75.4 | 59.7 | |
| COMBOODBackbone=RepVGG-A22026.03 | 78.6 | — | 60.8 | |
| D-KNNBackbone=RepVGG-A22026.03 | 78.1 | — | 62.6 | |
| CIDERBackbone=RepVGG-A22026.03 | 77.8 | — | 59.5 | |
| VDM log pT2026.05 | 77.7 | — | 60.5 | |
| ENN (EpiNet)Ensemble size=162024.05 | 77.4 | 73.8 | 63 | |
| Single NetworkEnsemble size=12024.05 | 75.6 | 71.5 | 64.7 | |
| EnergyBackbone=ResNet-502026.05 | 75.4 | — | 83.1 | |
| Score NetworkConfig=Hk = 4dk2026.05 | 74.6 | — | 79.6 | |
| L2EOOD score=Maximum Softmax Probability (MSP)2024.08 | 74.3 | — | — | |
| MSPBackbone=ResNet-502026.05 | 74.1 | — | 84.3 | |
| ViMBackbone=RepVGG-A22026.03 | 72.9 | — | 65.2 | |
| DEOOD score=Maximum Softmax Probability (MSP)2024.08 | 72.6 | — | — | |
| HMCOOD score=Maximum Softmax Probability (MSP)2024.08 | 72.5 | — | — | |
| CSGMCMCOOD score=Maximum Softmax Probability (MSP)2024.08 | 71.7 | — | — | |
| MahalanobisBackbone=RepVGG-A22026.03 | 71.4 | — | 72.3 | |
| E1 normed-forkBackbone=CLIP, Embedding Dimension=512-d2026.05 | 70.6 | — | 80 | |
| k-NNBackbone=CLIP, k-value=502026.05 | 69.9 | — | 78.6 | |
| k-NNBackbone=ResNet-50, k-value=502026.05 | 64.2 | — | 86.9 | |
| EigenScore2026.05 | 64.2 | — | — | |
| E3 normed-forkBackbone=ResNet-50, Embedding Dimension=2048-d2026.05 | 62.7 | — | 88.5 | |
| Proj. Regret2026.05 | 57.7 | — | — | |
| LMD2026.05 | 56.8 | — | — | |
| MC DropoutEnsemble size=162024.05 | 56.6 | 55.6 | 92.2 | |
| MahalanobisBackbone=ResNet-502026.05 | 48.3 | — | 98.1 | |
| DiffPath-6D2026.05 | 48.3 | — | — | |
| KNN (ℓ2)Backbone=RepVGG-A22026.03 | 19.7 | — | 98.7 |