I have described bad attempts to train RNN with Russian Traffic dataset before – here for example. To solve this problem first I have sorted all images into classes to it’s own directory and found that quality of images is very bad: 1. there is a lot of images which can’t recognize by human...