Table 5: Classification accuracy and OOD detection AUROC. ROT is the rotation augmentation; PRED means rot predictor; and IN-D means indistribution (CIFAR-10 in this experiment).
Method | ROT | PRED | JSD | Accuracy IN-D | OOD AUROC between CIFAR-10 and |
SVHN | LSUN | ImageNet | CIFAR-100 | nterp |
SimCLR(CSI[6] ) | ✓ | ✓ | - | - | 99.8 | 90.3 | 93.3 | 89.2 | 79.3 |
SimCLRs | - | - | - | 89.9 | 86.3 | 89.7 | 88.0 | 85.5 | 81.0 |
SimCLRs | ✓ | - | - | 92.2 | 97.4 | 91.0 | 91.5 | 90.1 | 82.1 |
SimCLRs | ✓ | ✓ | - | 90.9 | 98.9 | 89.7 | 92.8 | 89.5 | 79.3 |
SimCLRs(ours) | ✓ | ✓ | ✓ | 92.3 | 98.0 | 92.3 | 93.4 | 90.8 | 82.8 |
SimSiam | - | - | - | 91.4 | 91.4 | 91.7 | 90.5 | 87.7 | 81.9 |
SimSiam | ✓ | - | - | 91.9 | 97.4 | 91.0 | 91.4 | 89.8 | 82.9 |
SimSiam | ✓ | ✓ | - | 89.9 | 99.1 | 89.1 | 92.3 | 88.9 | 78.7 |
SimSiam(ours) | ✓ | ✓ | ✓ | 92.3 | 99.2 | 92.3 | 93.4 | 90.1 | 81.5 |