GEOMETRIC PRIOR-ASSISTED FEATURE PRESENTATION ENHANCEMENT FOR OBJECT DETECTION IN AERIAL IMAGES

Detecting objects in aerial images is an active yet challenging task due to the arbitrary orientation of aerial targets and the lack of details for small objects. Observing that geometric priors, e.g. some kinds of points and lines with specific geometrical properties, are helpful to determine the oriented bounding box, we proposed a Point-Line-Region (PLR) supervision module embedded in the FPN to learn the robust geometrical features, which are conducive to evaluate and locate the orientation, vertexes, and boundary of the bounding box. Our method improves feature representations of aerial targets by utilizing the supervision of across-category geometrical semantics, rather than their category semantics in the previous works. Extensive comparison experiments on the DOTA dataset prove that remarkable accuracy gains of mAP, about 2.1%, are achieved by integrating the PLR module in different architectures of networks.