Out-of-Distribution Detection on Average (OpenImage-O, Texture, iNaturalist, ImageNet-O)
96.23AUROCViM
Evaluation results
| Method | Links | ||
|---|---|---|---|
| ViMModel=ViT-B/16, Source=feat+logit2022.03 | 96.23 | 18.07 | |
| MahalanobisModel=ViT-B/16, Source=feat+label2022.03 | 96.02 | 19.45 | |
| ReActModel=ViT-B/16, Source=feat+logit2022.03 | 95.11 | 22.22 | |
| EnergyModel=ViT-B/16, Source=logit2022.03 | 94.9 | 22.43 | |
| ODINModel=ViT-B/16, Source=prob+grad2022.03 | 94.57 | 24.25 | |
| MaxLogitModel=ViT-B/16, Source=logit2022.03 | 94.57 | 24.25 | |
| ViMModel=Swin, Source=feat+logit2022.03 | 94.11 | 31.04 | |
| ResidualModel=ViT-B/16, Source=feat2022.03 | 92.93 | 30.23 | |
| ResidualModel=Swin, Source=feat2022.03 | 92.88 | 37.38 | |
| MahalanobisModel=Swin, Source=feat+label2022.03 | 92.16 | 40.39 | |
| ViMModel=BiT-S-R101x1, Source=feat+logit2022.03 | 90.91 | 41.46 | |
| KL MatchingModel=ViT-B/16, Source=prob2022.03 | 90.89 | 35.77 | |
| ReActModel=Swin, Source=feat2022.03 | 90.17 | 31.36 | |
| MSPModel=ViT-B/16, Source=prob2022.03 | 89.4 | 41.65 | |
| ViMModel=Res50d, Source=feat+logit2022.03 | 89.22 | 52.61 | |
| KL MatchingModel=Swin, Source=prob2022.03 | 88.87 | 46.99 | |
| MaxLogitModel=Swin, Source=logit2022.03 | 88.4 | 35.28 | |
| MahalanobisModel=Res50d, Source=feat+label2022.03 | 88.33 | 55.7 | |
| ODINModel=Swin, Source=prob+grad2022.03 | 88 | 36.58 | |
| ViMModel=RepVGG, Source=feat+logit2022.03 | 87.81 | 50.5 | |
| EnergyModel=Swin, Source=logit2022.03 | 87.77 | 35.08 | |
| MSPModel=Swin, Source=prob2022.03 | 87.57 | 43.44 | |
| ResidualModel=Res50d, Source=feat2022.03 | 87.01 | 58.55 | |
| MahalanobisModel=BiT-S-R101x1, Source=feat+label2022.03 | 86.62 | 53.34 | |
| MahalanobisModel=RepVGG, Source=feat+label2022.03 | 86.07 | 59.39 | |
| ViMModel=DeiT, Source=feat+logit2022.03 | 85.25 | 69.95 | |
| MahalanobisModel=DeiT, Source=feat+label2022.03 | 85.03 | 73.18 | |
| ReActModel=BiT-S-R101x1, Source=feat+logit2022.03 | 84.53 | 61.38 | |
| ResidualModel=RepVGG, Source=feat2022.03 | 84.19 | 59 | |
| ResidualModel=DeiT, Source=feat2022.03 | 84.15 | 74.13 | |
| ResidualModel=BiT-S-R101x1, Source=feat2022.03 | 84.14 | 56.23 | |
| KL MatchingModel=BiT-S-R101x1, Source=prob2022.03 | 83.63 | 55.62 | |
| KL MatchingModel=DeiT, Source=prob2022.03 | 83.49 | 64.8 | |
| ReActModel=Res50d, Source=feat2022.03 | 82.93 | 58.63 | |
| KL MatchingModel=Res50d, Source=prob2022.03 | 82.72 | 64.41 | |
| KL MatchingModel=RepVGG, Source=prob2022.03 | 81.35 | 61.65 | |
| MSPModel=DeiT, Source=prob2022.03 | 79.48 | 66.43 | |
| MaxLogitModel=BiT-S-R101x1, Source=logit2022.03 | 79.27 | 78.46 | |
| ODINModel=BiT-S-R101x1, Source=prob+grad2022.03 | 79.24 | 78.63 | |
| EnergyModel=BiT-S-R101x1, Source=logit2022.03 | 78.48 | 79.68 | |
| MSPModel=RepVGG, Source=prob2022.03 | 78.1 | 70.55 | |
| MSPModel=Res50d, Source=prob2022.03 | 77.99 | 67.96 | |
| ODINModel=RepVGG, Source=prob+grad2022.03 | 77.72 | 72.68 | |
| MaxLogitModel=RepVGG, Source=logit2022.03 | 77.56 | 73.5 | |
| ReActModel=DeiT, Source=feat2022.03 | 77.37 | 67 | |
| MSPModel=BiT-S-R101x1, Source=prob2022.03 | 77.25 | 77.83 | |
| ODINModel=DeiT, Source=prob+grad2022.03 | 77.13 | 63.92 | |
| MaxLogitModel=DeiT, Source=logit2022.03 | 76.79 | 64.49 | |
| EnergyModel=RepVGG, Source=logit2022.03 | 76.36 | 78.99 | |
| MaxLogitModel=Res50d, Source=logit2022.03 | 75.39 | 69.34 | |
| ODINModel=Res50d, Source=prob+grad2022.03 | 75.27 | 68.56 | |
| EnergyModel=DeiT, Source=logit2022.03 | 72.8 | 70.14 | |
| EnergyModel=Res50d, Source=logit2022.03 | 71.08 | 78.39 | |
| ReActModel=RepVGG, Source=feat2022.03 | 49.14 | 98.96 |