Multiple-dataset traffic sign classification with OneCNN

We take a look at current state of traffic sign classification discussing what makes it a specific problem of visual object classification. With impressive state-of-the-art results it is easy to forget that the domain extends beyond annotated datasets and overlook the problems that must be faced before we can start training classifiers. We discuss such problems, give an overview of previous work done, go over publicly available datasets and present a new one. Following that, classification experiments are conducted using a single CNN model, deeper than used previously and trained with dropout. We apply it over multiple datasets from Germany, Belgium and Croatia, their intersections and union, outperforming humans and other single CNN architectures for traffic sign classification.