The model architecture in this release is the same as the models in the veri776-pretrain repo
The differences are:
- center loss is added
- the images are cropped in the data augmentation process
| model | rank 1 |
|---|---|
| Resnet | 0.7460 |
| Resnext | 0.7707 |
| DenseNet | 0.7546 |
| SeResnet | 0.7658 |
| Swin Transformer | 0.7893 |