Train DBNet_r18 model in CTW1500 data set, the effect is relatively poor, can anyone help me see, is there any problem?

[qiyueliuhuo]

I use dbnet_resnet18_fpnc_100k_synthtext-2e9bf392.pth as pretrain model
The log file is as follows
20221221_222607.log

When training to 180 EPOCH, the model performance is the best. It seems that subsequent training is invalid. Why is this?

[qiyueliuhuo]

The result in the original paper is 82.4 76.6 79.4，My training results are 0.7635 0.6761 0.7171

[gaotongxiao]

We haven't trained DB on CTW yet, but one thing that I'm sure of is that the test-time image size should be different from IC15. While the authors didn't release their configuration for CTW, you could refer to other models' CTW configurations in MMOCR and try them on DB.

[qiyueliuhuo]

Thank you for your reply. There are indeed some post-processing parameters and test-image size are different from IC15. I am still trying...