The article can be found here
We choose a sample dataset from rvl-cdip with 10K trainig and 2k testing
| approach | accuracy | accuracy with finetuning | Finetune with res2net layer |
|---|---|---|---|
| VGG 19 (FC) | 61% | 87% | ? |
| VGG 19 (FV) | 74% | 75% | ? |
| VGG 19 (FV+FC) | 74% | 87% | ? |
| Resnet 152 (FC) | 67% | 69% | ? |
| Resnet 152 (FV) | 68% | 72% | ? |
| densenet161 (FV) | 68% | ? | ? |
| DFL | 92.6% | . | ? |