Authors: USTC-iFLYTEK

Description: 1. There are various shape of texts instance and enormous number of small objects in Large-scale Street View Text datasets. So we proposed method which combine the anchor-based method and segmentation-based method that can detects these text instance effectively.
2. For curve-shaped text instance, most scene text detection algorithms are deep learning based methods that depend on various scale of anchors and bounding box regression. But most existing quadrangular bounding box based detectors are difficult to locate texts with arbitrary shapes. So we use semantic segmentation methods to detect arbitrary shapes texts. But segmentation-based methods not worked well for detecting small objects. So we use anchor-based method to detect small objects, and use post processing to combine anchor-based and segmentation-based together.