AProNet: Detecting objects with precise orientation from aerial images

Abstract Detecting the arbitrary-oriented objects in aerial images is a significant yet challenging task in remote sensing image analysis. Existing methods generally use the oriented bounding boxes (OBBs) as Region of Interests (RoIs) to detect aerial objects, and learn the orientation angle information of objects under the supervision of OBB annotations. However, directly learning the orientation angle suffers from the training instability due to the existence of angular periodicity issue in object orientation prediction. In this study, we propose a novel axis projection-based angle learning network (termed as AProNet) for robust oriented object detection in aerial images. Instead of using the direct angle representation, we design an axis projection-based angle representation that is achieved by projecting the long axis of an aerial object along the X- and Y-axes in the image coordinate system. In this way, AProNet can obtain the orientation angle of objects based on the predicted axis projections, which are free of angular periodicity issue. Accordingly, a new loss function is developed to guide the training of AProNet. The loss function measures the loss between the predicted and groudtruth axis projections of objects and also dynamically balances the learning of the different OBB parameters. We also introduce a feature enhancement module to enhance the multi-scale features extracted by AProNet with geometric clues that are highly related to the axis projection-based angle learning. Extensive experiments demonstrate that the proposed axis projection-based angle learning can effectively handle the angular periodicity issue, achieving a competitive performance on two commonly-used aerial datasets DOTA and HRSC2016.