Oriented Object Detection with Fine-Grained Enhancement and Angle Constraint

With the development of aerial imaging technology, object detection in aerial images has attracted increasing attention. Although significant progress has been achieved, how to effectively perform feature extraction and boundary regression for oriented object detection remains a hard challenge. In this paper, a fine-grained enhanced detection network (FEDet) is proposed for oriented object detection in aerial images. Specifically, the bottom-up fine-grained enhancement module is designed to enrich the fine-grained information of features and optimize the detection performance. Furthermore, in order to guide the network to learn the angle information and achieve more accurate rotation angle prediction, the oriented-sensitive loss function is proposed with the embedded angle constraint. Extensive experiments on the widely used benchmarks (i.e., DOTA and HRSC2016) demonstrate that the proposed method achieves promising performance.