Affinity-Aware Relation Network for Oriented Object Detection in Aerial Images

Tingting Fang, Bin Liu, Zhiwei Zhao, Qi Chu, Nenghai Yu

Abstract: Object detection in aerial images is a challenging task due to the oriented and densely packed objects. However, densely packed objects constitute a significant characteristic of aerial images: objects are not randomly scattered around in images but in groups sharing similar orientations. Such a recurring pattern of object arrangement could enhance the rotated features and improve the detection performance. This paper proposes a novel and flexible Affinity-Aware Relation Network based on two-stage detectors. Specifically, an affinity-graph construction module is adopted to measure the affinity among objects and to select bounding boxes sharing high similarity with the reference box. Furthermore, we design a dynamic enhancement module, which uses the attention to learn neighbourhood message and dynamically determines weights for feature enhancement. Finally, we conduct experiments on several public benchmarks and achieve notable AP improvements as well as state-of-the-art performances on DOTA, HRSC2016 and UCAS-AOD datasets.