2p, Considering that scene texts are with arbitrary orientations, works in [22, 32, 46, 10, 26, 9] make the above methods possible for multi-oriented text detection. RRPN [32] introduces inclined anchors with angle information for arbitrary-oriented text prediction and rotated RoI pooling layer to project arbitrary-oriented proposals to the feature map for a text region classifier. TextBoxes++ [22] improves TextBoxes by regressing horizontal anchors to more general quadrilaterals enclosing oriented texts. It also proposes an efficient cascaded non-maximum suppression for quadrilaterals or rotated rectangles. With dense predictions and one step post processing, EAST [46] and DDR [10] both directly produce the rotated boxes or quadrangles of text ateach point in the text region. Recent text spotting methods like FOTS [26] and He et al. [9] show that training text detection and recognition simultaneously could greatly boost detection performance.