Domain Generalization on Terra Incognita (test)
86.02Location 38 cAccERM
Evaluation results
| Method | Links | ||||||||||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| ERMBackbone=ResNet-50, Pre-trained=ImageNet, Optimizer=Adam, Batch size=32, Learning rate=10^-4, Adversarial attack protocol=PGD (10 steps, epsilon=0.02)2026.03 | 86.02 | 9.82 | 54.86 | 23.59 | 80.35 | 16.37 | 56.44 | 38.42 | 79.06 | 3.61 | 43.43 | 15.02 | 92.34 | 12.13 | 54.5 | 26.32 | |
| IGNBackbone=ResNet-50, Pre-trained=ImageNet, Optimizer=Adam, Batch size=32, Learning rate=10^-4, Adversarial attack protocol=PGD (10 steps, epsilon=0.02)2026.03 | 85.87 | 36.03 | 54.89 | 21.49 | 81.41 | 13.07 | 60.69 | 37.64 | 79.59 | 3.5 | 48.25 | 11.95 | 91.82 | 8.44 | 49.41 | 16.49 | |
| DREBackbone=ResNet-50, Pre-trained=ImageNet, Optimizer=Adam, Batch size=32, Learning rate=10^-4, Adversarial attack protocol=PGD (10 steps, epsilon=0.02)2026.03 | 85.25 | 40.84 | 58.62 | 23.1 | 80.47 | 18.42 | 62.51 | 37.09 | 78.32 | 3.71 | 48.82 | 23.73 | 93.24 | 24.01 | 36.54 | 25.67 | |
| EGATBackbone=ResNet-50, Pre-trained=ImageNet, Optimizer=Adam, Batch size=32, Learning rate=10^-4, Adversarial attack protocol=PGD (10 steps, epsilon=0.02)2026.03 | 85.22 | 54.52 | 62.75 | 26.25 | 81.73 | 48.34 | 57.02 | 65.51 | 78.74 | 42.4 | 58.44 | 52.12 | 91.24 | 55.27 | 86.47 | 51.28 | |
| IGRBackbone=ResNet-50, Pre-trained=ImageNet, Optimizer=Adam, Batch size=32, Learning rate=10^-4, Adversarial attack protocol=PGD (10 steps, epsilon=0.02)2026.03 | 81.83 | 41.93 | 55.3 | 28.61 | 81.73 | 8.81 | 51.75 | 39.2 | 45.57 | 3.82 | 12.38 | 28.34 | 75.31 | 2.9 | 12.78 | 40.51 | |
| DMADABackbone=ResNet-50, Pre-trained=ImageNet, Optimizer=Adam, Batch size=32, Learning rate=10^-4, Adversarial attack protocol=PGD (10 steps, epsilon=0.02)2026.03 | 80.74 | 29.66 | 29.79 | 53.27 | 80.54 | 7.1 | 47.58 | 82.44 | 73.94 | 3.09 | 36.08 | 51.37 | 92.22 | 3.45 | 33.18 | 62.36 | |
| SGDropBackbone=ResNet-50, Pre-trained=ImageNet, Optimizer=Adam, Batch size=32, Learning rate=10^-4, Adversarial attack protocol=PGD (10 steps, epsilon=0.02)2026.03 | 77.8 | 27.87 | 39.35 | 46.11 | 78.57 | 10.23 | 38.02 | 74.08 | 76.47 | 3.29 | 36.05 | 49.62 | 91.19 | 0.52 | 38.93 | 67.71 | |
| IRMBackbone=ResNet-50, Pre-trained=ImageNet, Optimizer=Adam, Batch size=32, Learning rate=10^-4, Adversarial attack protocol=PGD (10 steps, epsilon=0.02)2026.03 | 65.77 | 20.18 | 30.04 | 34.86 | 52.59 | 9.73 | 19.98 | 36.18 | 60.08 | 3.83 | 20.63 | 35.05 | 83.93 | 4.55 | 7.61 | 35.85 |