Object Detection in Remote Sensing Images With Balanced Rotational and Horizontal Bounding Boxes

Object detection in remote sensing images is a challenging task because these images usually contain a number of targets with arbitrary orientations. To annotate the arbitrary-oriented objects accurately, rotational bounding boxes are more effective than horizontal bounding boxes. However, rotational bounding box deformation often happens when objects are near horizontal, since regressing angles is a highly nonlinear task. Aiming at this problem, we propose a region-based convolution neural network with balanced rotational and horizontal bounding boxes (RH-RCNN) for arbitrary-oriented object detection. We first design a multilayer-enhanced feature pyramid network (ML-FPN) to obtain powerful feature representations. Then we take both stability and accuracy into consideration, and devise an RH-head network to distinguish near horizontal objects from inclined ones. Angle prediction for near horizontal objects is prohibited to avoid large regression deviation. Rotational bounding boxes are only used to locate the obviously inclined objects. Experiments on a public large-scale dataset, named DOTA, show that not only the objective detection accuracy in terms of mAP but also the visualized predicted results are greatly improved.