How can I reproduce ICDAR13 acc?

[lironghua318]

When I train with ICDAR13 dataset with default config file, the train acc reach 1 quickly, but val acc is very low , which means overfitting .
So how can i reproduce the result on ICDAR2013?
thx

[lironghua318]

And pretrained model trained on which datasets?

[Pay20Y]

Directly only trained on ICDAR13 is not suitable, since there are too few images. You can trained on SynthText and Synth90K which can be found here. And you can finetune the model on ICDAR2013 but it is optional.

[lironghua318]

I will try the dataset,thx

[lironghua318]

I can't get 0.918 acc, just 0.889
Even when I test with your pretrained model, the ICDAR13 acc is 0.8928? not 0.918
can you offer more details?
my batch size is 64
i can reproduce SVTP and ICDAR2015

[lironghua318]

I finetune the model on ICDAR2013 and train for 625 iters and get 0.914acc

[Pay20Y]

I'm so sorry. I checked it again. The accuracy 91.8 is achieved on ICDAR13-857 not 1015.