Rotated Faster R-CNN for Oriented Object Detection in Aerial Images

Object detection have been widely used in the field of remote sensing. Different from natural scene images, the aerial images acquired by satellite and UAV are taken from birdview perspective. Common object detection algorithms suffer from the poor performance of detecting oriented targets. In this paper, we propose a Rotated Faster R-CNN to detect arbitrary oriented ground targets. On the basis of Faster R-CNN, we add a regression branch to predict the oriented bounding boxes for ground targets. Instead of removing the branch of predicting the horizontal bounding boxes, we train both two branches as a multi-task problem to improve the accuracy of our algorithms. And balanced FPN is used to improve the performance of detecting small targets in high resolution aerial images. We conduct experiments on DOTA dataset. Our methods could achieve competitive results of mAP 74.56 and FPS 13.0. The experiments prove that our algorithms show better results than previous algorithms in terms of accuracy.