An efficient and universal polygon prediction method based on derivable analytic geometry for arbitrary-shaped text detection

A polygon can represent the boundary of curved text more compactly than a rectangle. However, predicting reasonable polygon lacks of solutions due to the complex spatial relationships caused by having more vertices. The two main challenges are how to satisfy the constraints between vertices and how to cope with data conflicts caused by inconsistent annotation standards. To address these problems, we propose a divide and conquer methodology, in which a polygon is considered as a set of convex quadrangles. By predicting quadrangles in sequence, the vertices of the polygon are obtained consecutively and constrained by the previous ones. Then, we propose a measure for the overlap between convex quadrangles, with which the IoU between two polygons is calculated densely. Our method is derivable and can be trained end-to-end. Also, the polygon prediction branch that we proposed is universal and transplantable. We select basic architecture as the backbone, and the text/non-text classification branch adopts an online hard example mining strategy. Experiments on curved benchmark datasets, namely Total Text and CTW1500, demonstrate that our approach achieves state-of-the-art accuracy. It also maintains a high level of inferring efficiency.