The supplied models are trained on document-like images, so I wouldn't expect it to do particularly well on things like street signs. My experience with the new

The supplied models are trained on document-like images, so I wouldn't expect it to do particularly well on things like street signs. My experience with the new lstm based versions is that it's very much competitive with closed source solutions for document-like OCR.