Table 1.
All cross-validation results for different configurations. We consider same-sized cropping (SS) and random-resize cropping (RR) and different model input resolutions. Values are given in percent as mean and standard deviation over all five CV folds. Ensemble average refers to averaging over all predictions from all models. Ensemble optimal refers to averaging over the models we found with our search strategy for the optimal subset of configurations. C = 8 refers to training with eight classes without the unknown class. T1 refers to Task 1 without meta data and T2 refers to Task 2 with meta data. ResNext WSL 1 and 2 refer to ResNeXt-101 WSL 32 × 8d and 32 × 16d, respectively [8].
| Configuration | Sensitivity T1 | Sensitivity T2 |
|---|---|---|
| SENet154 SS 224 × 224 | 67.2 ± 0.8 | 70.0 ± 0.8 |
| ResNext WSL 1 SS 224 × 224 | 65.9 ± 1.6 | 68.1 ± 1.3 |
| ResNext WSL 2 SS 224 × 224 | 65.3 ± 0.8 | 69.1 ± 1.5 |
| EN B0 SS 224 × 224 C = 8 | 66.7 ± 1.8 | 68.8 ± 1.5 |
| EN B0 SS 224 × 224 | 65.8 ± 1.7 | 67.4 ± 1.6 |
| EN B0 RR 224 × 224 | 67.0 ± 1.6 | 68.9 ± 1.7 |
| EN B1 SS 240 × 240 | 65.9 ± 1.6 | 68.2 ± 1.8 |
| EN B1 RR 240 × 240 | 66.8 ± 1.5 | 68.5 ± 1.8 |
| EN B2 SS 260 × 260 | 67.2 ± 1.4 | 69.0 ± 2.5 |
| EN B2 RR 260 × 260 | 67.6 ± 2.0 | 70.1 ± 2.0 |
| EN B3 SS 300 × 300 | 67.8 ± 2.0 | 68.5 ± 1.7 |
| EN B3 RR 300 × 300 | 67.0 ± 1.5 | 68.4 ± 1.5 |
| EN B4 SS 380 × 380 | 67.8 ± 1.1 | 68.5 ± 1.1 |
| EN B4 RR 380 × 380 | 68.1 ± 1.6 | 69.4 ± 2.2 |
| EN B5 SS 456 × 456 | 68.2 ± 0.9 | 68.7 ± 1.6 |
| EN B5 RR 456 × 456 | 68.0 ± 2.2 | 69.0 ± 1.6 |
| EN B6 SS 528 × 528 | 68.8 ± 0.7 | 69.0 ± 1.4 |
| Ensemble Average | 71.7 ± 1.7 | 73.4 ± 1.6 |
| Ensemble Optimal | 72.5 ± 1.7 | 74.2 ± 1.1 |
| Official Testset | 63.6 | 63.4 |