Segmentation based rotated bounding boxes prediction and image synthesizing for object detection of high resolution aerial images

Abstract

Object detection for aerial images is becoming an active topic in computer vision with many real-world applications. It is a very challenging task due to many factors such as highly complex background, arbitrary object orientations, high input resolution, etc. In this paper, we develop a new training and inference mechanism, which is shown to significantly improve the detection accuracy for high resolution aerial images. Instead of estimating the orientations of objects using direct regressions like in previous methods, we propose to predict the rotated bounding boxes by leveraging a segmentation task, which is easier to train and yields more accurate detection results. In addition, an image synthesizing based data augmentation strategy is presented to address the data imbalance issues in aerial object detection. Extensive experiments have been conducted to verify our contribution. The proposed method sets new state-of-the-art performance on the challenging DOTA dataset. The source codes will be available at http://ice.dlut.edu.cn/lu/publications.html.