Out-of-Distribution Detection on ImageNet (ID) vs ImageNet-O Categories (OoD)
0.57OoD Detection: CastleSketched Lanczos Uncertainty
Evaluation Results
| Method | Links | ||||||||
|---|---|---|---|---|---|---|---|---|---|
| Sketched Lanczos UncertaintyModel Architecture=SWIN, Parameters (p)=200M, Memory Budget=3p2024.09 | 0.57 | 0.64 | 0.63 | 0.37 | 0.7 | 0.56 | 0.69 | 0.83 | |
| LLA-DModel Architecture=SWIN, Parameters (p)=200M, Memory Budget=3p2024.09 | 0.51 | 0.54 | 0.52 | 0.32 | 0.45 | 0.48 | 0.68 | 0.7 | |
| LEModel Architecture=SWIN, Parameters (p)=200M, Memory Budget=3p2024.09 | 0.51 | 0.58 | 0.62 | 0.41 | 0.59 | 0.47 | 0.67 | 0.8 | |
| LE-HModel Architecture=SWIN, Parameters (p)=200M, Memory Budget=3p2024.09 | 0.5 | 0.57 | 0.6 | 0.41 | 0.57 | 0.47 | 0.67 | 0.8 | |
| LLAModel Architecture=SWIN, Parameters (p)=200M, Memory Budget=3p2024.09 | 0.49 | 0.56 | 0.59 | 0.41 | 0.56 | 0.45 | 0.67 | 0.78 | |
| SWAGModel Architecture=SWIN, Parameters (p)=200M, Memory Budget=3p2024.09 | 0.46 | 0.54 | 0.55 | 0.52 | 0.44 | 0.36 | 0.64 | 0.72 |