4.2.1. ICDAR2015

This is a standard dataset proposed for scene text detection in the Challenge4 of ICDAR2015 Robust Reading Competition, which is divided into two categories: the training part contains 1000 image-text pairs; the testing part contains 500 image-text pairs. Each picture is associated with one or more labels annotated with four vertices of the quadrangle. Unlike the previous datasets (such as ICDAR2013 [37]) that only contain horizontal text, the orientations of the reference text in this benchmark are arbitrary.
