CODR-Det: Rotated Object Detection Based on Circumcircle Fixed Point Offset and Dynamic Rotation Perception in Remote Sensing Images

Compared with natural senses, object detection methods in remote sensing images encounter some challenges due to the deviation of feature extraction and inaccurate positioning caused by the characteristics of remote sensing images, such as arbitrary direction, complex background and diverse shapes. To address these problems, we propose a rotated object detection method based on Circumcircle fixed point Offset and Dynamic Rotation perception (CODR-Det). This method consists of two parts: the representation of fixed point offset rotation frame based on Circumcircle fixed point Offset (CFPO) and Dynamic Rotation Perception Detection Head (DRP Head). First, according to the geometric relationship, CFPO obtains the unique rotated prediction box by predicting the offsets of two corner points of the rotated bounding box relative to the fixed point, which is on the circumscribed circle of the rotated bounding box. This representation method can generate the rotated bounding box without complex post-processing, and avoid the occurrence of boundary problems. Second, DRP Head enables the model to adaptively extract the features of the object according to the characteristics of the object by introducing Omni-dimensional Dynamic Dilated Convolution (ODDConv) layer and Adaptive Rotated Dilated Convolution (ARDConv) layer. The results on the HRSC2016 and DOTA datasets show that CODR-Det achieves 90.60% and 76.31% detection accuracy, respectively, which indicates better performance than other methods.