Out-of-Distribution Detection on ImageNet (vs. Places/SUN/iNaturalist/Textures)
96.2Average AUROCCFS(1×2)
Evaluation results
| Method | Links | ||||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|
| CFS(1×2)Cost=1F2026.05 | 96.2 | — | — | — | — | — | — | — | — | — | |
| DICEBackbone=ResNetV2-50, Pre-training=BiT-S2025.05 | 91.72 | 38.55 | 89.18 | 29.67 | 92.15 | 8.46 | 97.73 | 40.77 | 87.81 | 29.36 | |
| GradOrthBackbone=ResNetV2-50, Pre-training=BiT-S2025.05 | 90.34 | 55.24 | 82.914 | 43.67 | 88.28 | 19.85 | 95.122 | 21.112 | 95.03 | 34.97 | |
| GradPCABackbone=ResNetV2-50, Pre-training=BiT-S2025.05 | 90.22 | 61.27 | 81.75 | 49.35 | 87.2 | 17.78 | 95.8 | 18.08 | 96.14 | 36.62 | |
| GradPCA+DICEBackbone=ResNetV2-50, Pre-training=BiT-S2025.05 | 89.41 | 65.07 | 80.1 | 53.16 | 85.85 | 19.25 | 95.48 | 18.16 | 96.2 | 38.91 | |
| GAIA-ABackbone=ResNetV2-50, Pre-training=BiT-S2025.05 | 87.85 | 53.54 | 85.66 | 46.05 | 89.33 | 21.69 | 95.33 | 69.78 | 81.07 | 47.77 | |
| Revisited PCABackbone=ResNetV2-50, Pre-training=BiT-S2025.05 | 87.36 | 64.42 | 80.54 | 57.32 | 83.68 | 20.05 | 95.33 | 35.55 | 89.91 | 44.83 | |
| GradPCA+DICEBackbone=ResNetV2-101, Pre-training=BiT-M2025.05 | 85.49 | 76.97 | 75.51 | 64.82 | 82.95 | 37.68 | 93.1 | 44.07 | 90.4 | 55.88 | |
| DiffPath-6DCost=10F2026.05 | 85 | — | — | — | — | — | — | — | — | — | |
| EnergyBackbone=ResNetV2-50, Pre-training=BiT-S2025.05 | 83.85 | 70.79 | 78.91 | 66.36 | 81.28 | 25.21 | 94.38 | 64.2 | 80.84 | 56.64 | |
| Max logitsBackbone=ResNetV2-50, Pre-training=BiT-S2025.05 | 82.66 | 73.73 | 77.77 | 69.78 | 79.92 | 29.76 | 93.48 | 68.96 | 79.47 | 60.56 | |
| ODINBackbone=ResNetV2-50, Pre-training=BiT-S2025.05 | 82.59 | 73.86 | 77.7 | 70.03 | 79.84 | 29.98 | 93.43 | 69.18 | 79.4 | 60.76 | |
| ReActBackbone=ResNetV2-50, Pre-training=BiT-S2025.05 | 82.48 | 69.42 | 76.41 | 63.95 | 78.68 | 24.83 | 93.45 | 58.81 | 81.36 | 54.25 | |
| GradPCABackbone=ResNetV2-101, Pre-training=BiT-M2025.05 | 82.1 | 83.12 | 71.55 | 75.16 | 78.68 | 45.84 | 91.69 | 60.74 | 86.49 | 66.22 | |
| Kernel PCA (CoRP)Backbone=ResNetV2-101, Pre-training=BiT-M2025.05 | 78.77 | 86.61 | 60.37 | 83.81 | 64.63 | 25.03 | 94.43 | 28.82 | 93.65 | 56.07 | |
| Kernel PCA (CoRP)Backbone=ResNetV2-50, Pre-training=BiT-S2025.05 | 78.65 | 81.83 | 65.62 | 76.97 | 71.21 | 66.3 | 81.31 | 15.37 | 96.48 | 60.11 | |
| Proj. GradsBackbone=ResNetV2-50, Pre-training=BiT-S2025.05 | 78.03 | 71.04 | 75.73 | 66.35 | 78.56 | 32.52 | 89.71 | 73.2 | 68.1 | 60.78 | |
| NCI (w/o filter)Backbone=ResNetV2-50, Pre-training=BiT-S2025.05 | 77.25 | 91.14 | 68.52 | 88.61 | 72.93 | 69.62 | 84.63 | 65.91 | 82.9 | 78.82 | |
| Revisited PCABackbone=ResNetV2-101, Pre-training=BiT-M2025.05 | 74.89 | 80.76 | 64.12 | 75.78 | 67.98 | 27.89 | 92.11 | 73.3 | 74.69 | 64.43 | |
| MahalanobisBackbone=ResNetV2-101, Pre-training=BiT-M2025.05 | 74.72 | 91.82 | 56.75 | 92.03 | 57.58 | 33.01 | 92.07 | 31 | 92.48 | 61.97 | |
| NCI (w/o filter)Backbone=ResNetV2-101, Pre-training=BiT-M2025.05 | 74.37 | 92.41 | 63.75 | 92.41 | 69.69 | 46.84 | 91.5 | 86.36 | 72.52 | 79.5 | |
| Proj. GradsBackbone=ResNetV2-101, Pre-training=BiT-M2025.05 | 73.55 | 84.02 | 72.48 | 82.31 | 75.4 | 57.51 | 83.69 | 91.32 | 62.63 | 78.29 | |
| KNNBackbone=ResNetV2-101, Pre-training=BiT-M2025.05 | 72.97 | 90.33 | 56.29 | 92.14 | 57.22 | 37.95 | 90.85 | 47.49 | 89.78 | 67.15 | |
| GAIA-ZBackbone=ResNetV2-50, Pre-training=BiT-S2025.05 | 72.72 | 94.77 | 58.63 | 93.69 | 62.28 | 85.23 | 75.82 | 26.44 | 94.17 | 75.03 | |
| MSPBackbone=ResNetV2-50, Pre-training=BiT-S2025.05 | 71.51 | 84.4 | 67.28 | 82.79 | 67.94 | 52.62 | 83.94 | 83.04 | 66.88 | 75.71 | |
| DICEBackbone=ResNetV2-101, Pre-training=BiT-M2025.05 | 70.58 | 88.34 | 65.87 | 81.67 | 71.39 | 71.53 | 87.4 | 96 | 57.65 | 84.38 | |
| GAIA-ZBackbone=ResNetV2-101, Pre-training=BiT-M2025.05 | 68.71 | 97.81 | 56.09 | 98.77 | 57.51 | 79.55 | 79.15 | 62.39 | 82.08 | 84.63 | |
| EnergyBackbone=ResNetV2-101, Pre-training=BiT-M2025.05 | 67.6 | 86.94 | 60.96 | 83.94 | 63.36 | 46.07 | 87.17 | 94.26 | 58.91 | 77.8 | |
| Max logitsBackbone=ResNetV2-101, Pre-training=BiT-M2025.05 | 66.28 | 89.36 | 59.98 | 87.94 | 61.69 | 55.73 | 84.78 | 94.85 | 58.66 | 81.97 | |
| ODINBackbone=ResNetV2-101, Pre-training=BiT-M2025.05 | 66.28 | 89.37 | 59.98 | 87.95 | 61.68 | 55.74 | 84.78 | 94.85 | 58.66 | 81.98 | |
| GradOrthBackbone=ResNetV2-101, Pre-training=BiT-M2025.05 | 66.08 | 95.45 | 66.006 | 93.52 | 71.32 | 96.68 | 67.57 | 97.857 | 59.44 | 95.88 | |
| MSPBackbone=ResNetV2-101, Pre-training=BiT-M2025.05 | 61.74 | 92.84 | 57.05 | 92.97 | 56.72 | 71.28 | 76.5 | 95.44 | 56.71 | 88.13 | |
| ReActBackbone=ResNetV2-101, Pre-training=BiT-M2025.05 | 61.36 | 92.86 | 53.85 | 91.21 | 55.73 | 62.24 | 83.02 | 96.44 | 52.85 | 85.68 | |
| KNNBackbone=ResNetV2-50, Pre-training=BiT-S2025.05 | 61.31 | 92.68 | 51.01 | 92.96 | 51.78 | 93.69 | 48.81 | 24.12 | 93.64 | 75.86 | |
| GAIA-ABackbone=ResNetV2-101, Pre-training=BiT-M2025.05 | 47.42 | 95.77 | 47.23 | 94.65 | 51.41 | 85.31 | 60.07 | 98.95 | 30.96 | 93.67 | |
| MahalanobisBackbone=ResNetV2-50, Pre-training=BiT-S2025.05 | 43.59 | 97.7 | 30.61 | 97.56 | 28.53 | 99.09 | 23.3 | 30.21 | 91.93 | 81.14 |