The Unified Object Detection Framework with Arbitrary Angle

Abstract

Traditional horizontal bounding boxes are hard to locate targets with various directions. In the complex background, this method also cannot distinguish object from background very well and it is difficult to separate dense objects. Therefore, the traditional deep learning model is far from meeting the need for object detection in arbitrary orientation scenarios. In order to solve these problems, the unified object detection framework with arbitrary orientation is proposed in this paper. The angle is used as a variable for regression, which can accurately detect the rotating object, its position and direction information in complex background. The rotation region of interest (RRoI) align is proposed to replace region of interest(RoI) pooling, which can improve the precision of the detection model. Besides, a skew intersection over union (IoU) calculation method is adopted to solve the problem that densely placed objects are hard to detect and improve the recall of the detection model, which can reduce redundant detection regions. Experiments on remote sensing data set show that our method has better robustness and performance than other algorithms in detecting small targets. Compared with other algorithms, our algorithm is not only applicable to ship detection, but also to remote sensing images, document detection and other arbitrary-oriented object detection scenarios.