Oriented Object Detection Based on CNN and Transformer for Remote Sensing Images

Remote sensing image object detection has broad application prospects in civil fields, such as environmental monitoring and resource exploration. Since remote sensing images are characterized by complex and diverse imaging scenes, there is no generalized and universal remote sensing image object detection method yet. This paper proposes an anchorless fixed-frame remote sensing image object detection method using CNN and transformer, and the detection head based on centroid localization and boundary-aware vectors. The method is a single-stage algorithm, which does not depend on the object candidate frame and avoids the complicated anchor frame design and tedious anchor frame parameter tuning process. Experiments were conducted on the object detection dataset DOTA 2.0, and the results show the superiority of our proposed method when compared with state-of-the-art methods.