Results in the paper

[janzd]

Hi!
I've been trying to reproduce the results in your paper on IAM dataset by generating new data with the code in this repository, and using the model for text recognition that you talk about and quote in the paper as [3] https://github.com/clovaai/deep-text-recognition-benchmark/tree/1863c76b6a2e672373894534a8f0406f122dd5b3
However, when I train the recognition model (using the best architecture as you say in the paper) on IAM, the WER is about 20% without any data augmentation, which is 5% better than you state in the paper. Better performance wouldn't really be a problem, but the main issue I have is that I can't achieve any improvement in performance by augmenting the dataset with new data samples generated by your model.
I use the default settings that are set in the options files and stated in the readme of this repository. Could you provide detailed settings for training of your proposed model and for training of the text recognition model that you used, with which you were able to achieve the results that you state in your paper?

[rlit]

Hi @kurapan .
I am afraid that the exact conditions for reproducing these experiments are a little hard for us to recall, as we did not keep the exact hyper-parameters we used for clova.