Authors: Yanbing Dong ( from MetaSota.ai)

Description: There are horizontal, oriented or curved text in this dataset, so axis-aligned rectangles, rotated rectangles or quadrangles are not enough to represent these text instances, here we use Mask RCNN to get a mask to represent a text instance, then get a polygon from the mask as a representation of the text in-stance. At last, do NMS based on overlap of different text instance instead of IOU, because a long text instance can be detected to one instance or some separated text instances at the same time.