Out-of-Distribution Detection on CIFAR-10
100AUROCPixelCNN++
Evaluation Results
| Method | Links | ||||||||
|---|---|---|---|---|---|---|---|---|---|
| PixelCNN++score_type=Bits Per Pixel (BPP), test_distribution=Gaussian2018.12 | 100 | 0 | 100 | — | — | — | — | — | |
| PixelCNN++ + OEscore_type=Bits Per Pixel (BPP), test_distribution=Gaussian2018.12 | 100 | 0 | 99.6 | — | — | — | — | — | |
| PixelCNN++ + OEscore_type=Bits Per Pixel (BPP), test_distribution=Blobs2018.12 | 99.5 | 1.3 | 96.2 | — | — | — | — | — | |
| Auxiliary Rotation Predictionaverage_of_runs=52019.06 | 99 | 1.2 | 85.6 | — | — | — | — | — | |
| DOEApproach=Fine-tuning, Backbone=WRN-40-22023.03 | 98.78 | 5.15 | — | — | — | — | — | — | |
| OEApproach=Fine-tuning, Backbone=WRN-40-22023.03 | 97.85 | 12.41 | — | — | — | — | — | — | |
| MSPOOD Detector=Maximum Softmax Probability, Backbone=Wide Residual Network, Outlier Exposure=true2018.12 | 97.8 | 9.5 | 90.5 | — | — | — | — | — | |
| OE+UMAuxiliary Outliers=true2023.06 | 97.6 | 11.22 | 97.87 | — | 93.66 | — | — | — | |
| OE+UMAuxiliary Data=true2023.06 | 97.6 | 11.22 | 97.87 | — | 93.66 | — | — | — | |
| MixOEApproach=Fine-tuning, Backbone=WRN-40-22023.03 | 97.59 | 13.55 | — | — | — | — | — | — | |
| KNN+Backbone=DenseNet-1012022.04 | 97.58 | 12.16 | — | — | 94.45 | — | — | — | |
| DOE+UMauxiliary_outliers=true2023.06 | 97.49 | 11.62 | 97.63 | — | 90.57 | — | — | — | |
| OE+UMAPAuxiliary Outliers=true2023.06 | 97.48 | 12.21 | 97.74 | — | 94.01 | — | — | — | |
| DOE+UMAPauxiliary_outliers=true2023.06 | 97.48 | 11.85 | 96.62 | — | 91.02 | — | — | — | |
| OE+UMAPAuxiliary Data=true2023.06 | 97.48 | 12.21 | 97.74 | — | 93.44 | — | — | — | |
| DOEauxiliary_outliers=true2023.06 | 97.41 | 12.84 | 97.38 | — | 91.02 | — | — | — | |
| SSD+Approach=Fine-tuning, Backbone=WRN-40-22023.03 | 97.36 | 14.84 | — | — | — | — | — | — | |
| WOODSauxiliary_outliers=true2023.06 | 97.32 | 12.94 | 96.78 | — | 92.22 | — | — | — | |
| OEAuxiliary Outliers=true2023.06 | 97.07 | 13.8 | 97.31 | — | 92.59 | — | — | — | |
| OEAuxiliary Data=true2023.06 | 97.07 | 13.8 | 97.31 | — | 92.59 | — | — | — | |
| SSD+Backbone=DenseNet-1012022.04 | 96.96 | 16.21 | — | — | 94.45 | — | — | — | |
| OEwatermarking=true2022.10 | 96.89 | 19.06 | 99.31 | — | — | — | — | — | |
| CSIApproach=Fine-tuning, Backbone=WRN-40-22023.03 | 96.87 | 17.39 | — | — | — | — | — | — | |
| OEwatermarking=false2022.10 | 96.41 | 20.96 | 99.39 | — | — | — | — | — | |
| Maximum Softmax Probability (MSP)average_of_runs=52019.06 | 96.3 | 8.1 | 70.8 | — | — | — | — | — | |
| Energy+UMAPAuxiliary Data=true2023.06 | 95.63 | 17.51 | 95.92 | — | 88.12 | — | — | — | |
| ASH-S+UMauxiliary_outliers=false2023.06 | 95.22 | 23.94 | 95.53 | — | 92.47 | — | — | — | |
| Energywatermarking=true2022.10 | 95.21 | 23.69 | 98.83 | — | — | — | — | — | |
| ASH-Sauxiliary_outliers=false2023.06 | 95.13 | 25.87 | 95.51 | — | 93.69 | — | — | — | |
| EnergyAuxiliary Outliers=true2023.06 | 94.58 | 18.79 | 94.69 | — | 80.91 | — | — | — | |
| EnergyAuxiliary Data=true2023.06 | 94.58 | 18.79 | 94.69 | — | 80.91 | — | — | — | |
| ASH-S+UMAPauxiliary_outliers=false2023.06 | 94.57 | 25.95 | 94.86 | — | 93.69 | — | — | — | |
| POEMAuxiliary Outliers=true2023.06 | 94.37 | 18.5 | 94.51 | — | 77.24 | — | — | — | |
| POEMAuxiliary Data=true2023.06 | 94.37 | 18.5 | 94.51 | — | 77.24 | — | — | — | |
| POEM+UMAPAuxiliary Data=true2023.06 | 94.18 | 20.55 | 94.15 | — | 76.62 | — | — | — | |
| Energy+UMAPAuxiliary Outliers=false2023.06 | 93.97 | 30.71 | 94.38 | — | 94.01 | — | — | — | |
| Energy+UMAPAuxiliary Data=false2023.06 | 93.97 | 30.71 | 94.38 | — | 94.01 | — | — | — | |
| ODINBackbone=DenseNet-1012022.04 | 93.86 | 30.02 | — | — | 94.38 | — | — | — | |
| Energy+UMAuxiliary Outliers=false2023.06 | 93.73 | 33.29 | 94.27 | — | 92.8 | — | — | — | |
| Energy+UMAuxiliary Data=false2023.06 | 93.73 | 33.29 | 94.27 | — | 92.8 | — | — | — | |
| ODINwatermarking=true2022.10 | 93.5 | 37.06 | 98.55 | — | — | — | — | — | |
| KNNApproach=Post-hoc, Backbone=WRN-40-22023.03 | 93.43 | 34.56 | — | — | — | — | — | — | |
| Mahalanobiswatermarking=false2022.10 | 93.23 | 34.18 | 98.41 | — | — | — | — | — | |
| MahalanobisApproach=Post-hoc, Backbone=WRN-40-22023.03 | 93.21 | 34.98 | — | — | — | — | — | — | |
| PixelCNN++score_type=Bits Per Pixel (BPP), test_distribution=Blobs2018.12 | 93.2 | 17.2 | 60 | — | — | — | — | — | |
| POEM+UMAuxiliary Data=true2023.06 | 93.04 | 23.52 | 92.99 | — | 67.41 | — | — | — | |
| Energy+UMAuxiliary Data=true2023.06 | 93.02 | 24.41 | 92.36 | — | 71.97 | — | — | — | |
| ReActwatermarking=true2022.10 | 93.01 | 38.91 | 98.42 | — | — | — | — | — | |
| Softmaxwatermarking=true2022.10 | 92.91 | 42.86 | 98.42 | — | — | — | — | — | |
| ReActauxiliary_outliers=false2023.06 | 92.76 | 38.43 | 93.57 | — | 93.74 | — | — | — | |
| GODINBackbone=DenseNet-1012022.04 | 92.48 | 28.98 | — | — | 94.22 | — | — | — | |
| EnergyBackbone=DenseNet-1012022.04 | 92.44 | 30.16 | — | — | 94.38 | — | — | — | |
| EnergyAuxiliary Outliers=false2023.06 | 92.07 | 42.69 | 92.72 | — | 94.01 | — | — | — | |
| EnergyAuxiliary Data=false2023.06 | 92.07 | 42.69 | 92.72 | — | 94.01 | — | — | — | |
| MSPBackbone=DenseNet-1012022.04 | 92.05 | 49.95 | — | — | 94.38 | — | — | — | |
| ETNevaluation_mode=MP2026.04 | 91.8 | — | — | — | — | — | — | — | |
| Energy+UMBackbone=WRN-40-42023.06 | 91.74 | 40.4 | 92.67 | — | 92.68 | — | — | — | |
| VOSApproach=Fine-tuning, Backbone=WRN-40-22023.03 | 91.56 | 31.55 | — | — | — | — | — | — | |
| ODINAuxiliary Outliers=false2023.06 | 91.46 | 42.31 | 91.67 | — | 94.01 | — | — | — | |
| ODINAuxiliary Data=false2023.06 | 91.46 | 42.31 | 91.67 | — | 94.01 | — | — | — | |
| PixelCNN++ + OEscore_type=Bits Per Pixel (BPP), test_distribution=LSUN2018.12 | 90.9 | 43.2 | 72.4 | — | — | — | — | — | |
| DeepEnsevaluation_mode=MP2026.04 | 90.78 | — | — | — | — | — | — | — | |
| DICEauxiliary_outliers=false2023.06 | 90.66 | 41.51 | 91.11 | — | 92.55 | — | — | — | |
| DMMevaluation_mode=MP2026.04 | 90.65 | — | — | — | — | — | — | — | |
| Energywatermarking=false2022.10 | 90.56 | 37.67 | 97.46 | — | — | — | — | — | |
| LAevaluation_mode=MP2026.04 | 90.23 | — | — | — | — | — | — | — | |
| ODINwatermarking=false2022.10 | 90.21 | 42.64 | 97.49 | — | — | — | — | — | |
| GradNormwatermarking=false2022.10 | 90.1 | 40.51 | 97.35 | — | — | — | — | — | |
| MSPAuxiliary Outliers=false2023.06 | 89.9 | 60.08 | 91.48 | — | 94.01 | — | — | — | |
| MSPAuxiliary Data=false2023.06 | 89.9 | 60.08 | 91.48 | — | 94.01 | — | — | — | |
| Softmaxwatermarking=false2022.10 | 89.82 | 55.7 | 97.32 | — | — | — | — | — | |
| PixelCNN++ + OEscore_type=Bits Per Pixel (BPP), test_distribution=Places3652018.12 | 89.3 | 50.3 | 70.4 | — | — | — | — | — | |
| MSPOOD Detector=Maximum Softmax Probability, Backbone=Wide Residual Network, Outlier Exposure=false2018.12 | 89.3 | 34.9 | 59.2 | — | — | — | — | — | |
| Energy+UMAPBackbone=WRN-40-42023.06 | 88.84 | 50.23 | 89.31 | — | 93.86 | — | — | — | |
| PixelCNN++ + OEscore_type=Bits Per Pixel (BPP), test_distribution=Textures2018.12 | 88.8 | 48.9 | 70 | — | — | — | — | — | |
| ReActwatermarking=false2022.10 | 88.79 | 57.59 | 97.37 | — | — | — | — | — | |
| ODINApproach=Post-hoc, Backbone=WRN-40-22023.03 | 88.69 | 42.8 | — | — | — | — | — | — | |
| MSPApproach=Post-hoc, Backbone=WRN-40-22023.03 | 88.4 | 53.77 | — | — | — | — | — | — | |
| Free EnergyApproach=Post-hoc, Backbone=WRN-40-22023.03 | 88.27 | 37.77 | — | — | — | — | — | — | |
| ETNevaluation_mode=UM2026.04 | 87.98 | — | — | — | — | — | — | — | |
| MAP_EDLevaluation_mode=MP2026.04 | 87.91 | — | — | — | — | — | — | — | |
| MAP_CEevaluation_mode=MP2026.04 | 87.71 | — | — | — | — | — | — | — | |
| EnergyBackbone=WRN-40-42023.06 | 87.69 | 58.47 | 88.16 | — | 93.86 | — | — | — | |
| IB-EDLevaluation_mode=MP2026.04 | 87.58 | — | — | — | — | — | — | — | |
| MahalanobisBackbone=DenseNet-1012022.04 | 87.56 | 35.88 | — | — | 94.38 | — | — | — | |
| MSPBackbone=WRN-40-42023.06 | 87.12 | 68.29 | 87.84 | — | 93.86 | — | — | — | |
| MAP_EDLevaluation_mode=UM2026.04 | 86.98 | — | — | — | — | — | — | — | |
| PEFT‡Pre-trained model=OpenCLIP2025.09 | 86.39 | 57.38 | — | — | — | 86.95 | 85.38 | — | |
| IB-EDLevaluation_mode=UM2026.04 | 86.19 | — | — | — | — | — | — | — | |
| PEFT†Pre-trained model=CLIP2025.09 | 83.97 | 61.98 | — | — | — | 84.42 | 82.61 | — | |
| Deep ens.2020.07 | 83.9 | — | — | — | — | — | — | — | |
| PixelCNN++ + OEscore_type=Bits Per Pixel (BPP), test_distribution=Mean2018.12 | 83.7 | 46.4 | 66 | — | — | — | — | — | |
| Deep ens. (LPBN)2020.07 | 83.4 | — | — | — | — | — | — | — | |
| DMMevaluation_mode=UM2026.04 | 83.36 | — | — | — | — | — | — | — | |
| ODINBackbone=WRN-40-42023.06 | 83.29 | 65.68 | 82.74 | — | 93.86 | — | — | — | |
| BatchEnsemble2020.07 | 82.9 | — | — | — | — | — | — | — | |
| LPBNaveraging=false2020.07 | 82.7 | — | — | — | — | — | — | — | |
| MC-Dropout2020.07 | 82.4 | — | — | — | — | — | — | — | |
| ReActApproach=Post-hoc, Backbone=WRN-40-22023.03 | 82.21 | 58.22 | — | — | — | — | — | — | |
| Dropoutmode=Mean2020.07 | 82.2 | — | — | — | — | — | — | — |