Rotated cascade R-CNN: A shape robust detector with coordinate regression

General object detection task mainly takes axis-aligned bounding-boxes as the detection outputs. To address more challenging scenarios, such as curved text detection and multi-oriented object detection in aerial images, we propose a novel two-stage approach for shape robust object detection. In the first stage, a locally sliding line-based point regression (LocSLPR) approach is presented to estimate the outline of the object, which is denoted as the intersections of the sliding lines and the bounding-box of the object. To make full use of information, we only regress partial coordinates and calculate the remaining coordinates by the sliding rule. We find that regression can achieve higher precision with fewer parameters than the segmentation method. In the second stage, a rotated cascade region-based convolutional neural network (RCR-CNN) is used to gradually regress the target object, which can further improve the performance of our system. Experiments demonstrate that our method achieves state-of-the-art performance in several quadrangular object detection tasks. For example, our method yielded a score of 0.796 in the ICPR 2018 Contest on Robust Reading for Multi-Type Web Images, where we won first place for text detection tasks. The method also achieved 69.2% mAP on Task 1 of the ICPR 2018 Contest on Object Detection in Aerial Images, which was our best single model, where we also won first place. In addition, the method outperforms the previously published best record on the curved text dataset (CTWI500). (C) 2019 Elsevier Ltd. All rights reserved.