Table 1 Typical methods and elapsed times for each of the top models in the competition
1st place model | 2nd place model | 3rd place model | 4th place model | Matsuoka et al. (2018) | |
|---|---|---|---|---|---|
CNN architecture | PyramidNet | WideResNet | Shake-shake ResNet26 | MobileNetV2 | LeNet |
Number of parameters | 7.6 M | 4.3 M | 3.0 M | 11.6 M | 0.60 M |
Preprocessing | Binarization | – | Downsampling (32 × 32) | Upsampling (96 × 96) | – |
Oversampling (Data augmentation) | Vertical flip Horizontal flip Cutout Random shift | Cropping Random rotation Random erasing | Random crop + Padding Horizontal flip | Random rotation Mixup | – |
Undersampling | Hard negative mining | Random sampling | Random sampling | Random sampling | Random sampling |
Ensemble learning | 5 models (Different hard negative ratio) | – | – | 5 models (Different learning rate/preproccesing) | 10 models (Different negative samples) |
TTA | Same as training phase | 5 Crop × 4 Rotation | 10 Crop + Padding | – | – |
Others | – | Focal loss | RReLU | – | – |
Time for training | 10 days | 1 day | 4 days | 5 days | 15 h |
Time for test | 10 h | 1 h | 1 h | 5 h | 4 h |