Vector Decomposition-Based Arbitrary-Oriented Object Detection for Optical Remote Sensing Images

Arbitrarily oriented object detection is one of the most-popular research fields in remote sensing image processing. In this paper, we propose an approach to predict object angles indirectly, thereby avoiding issues related to angular periodicity and boundary discontinuity. Our method involves representing the long edge and angle of an object as a vector, which we then decompose into horizontal and vertical components. By predicting the two components of the vector, we can obtain the angle information of the object indirectly. To facilitate the transformation between angle-based representation and the proposed vector-decomposition-based representation, we introduced two novel techniques: angle-to-vector encode (ATVEncode) and vector-to-angle decode (VTADecode). These techniques not only improve the efficiency of data processing, but also accelerate the training process. Furthermore, we propose an adaptive coarse-to-fine positive-negative-sample-selection (AdaCFPS) method based on the vector-decomposition-based representation of the object. This method utilizes the Kullback-Leibler divergence loss as a matching degree to dynamically select the most-suitable positive samples. Finally, we modified the YOLOX model to transform it into an arbitrarily oriented object detector that aligns with our proposed vector-decomposition-based representation and positive-negative-sample-selection method. We refer to this redesigned model as the vector-decomposition-based object detector (VODet). In our experiments on the HRSC2016, DIOR-R, and DOTA datasets, VODet demonstrated notable advantages, including fewer parameters, faster processing speed, and higher precision. These results highlighted the significant potential of VODet in the context of arbitrarily oriented object detection.