TDC560

TDC560 dataset contains 560 difficult images, one of which are selected from the testing set of CTW1500 and TD500, others are generated by ourselves with text-line annotations. In the selecting process, we sort images with the extreme spatial distances between characters and words. Additionally, to bridge hard text-line detection to real world, we rich existing diverse image sources with our own data, which has two significant merits: (1)we expand abundant images containing Chinese texts which is relatively lack in previous benchmarks such as CTW1500; (2) we collect various types of images such as scene text, design text and some hard stylish text. Some visualization results are demonstrated in the paper.