PseKD: Phase-Shift Encoded Knowledge Distillation for Oriented Object Detection in Remote Sensing Images

Chao Wang, Yubiao Yue, Bingchun Luo, Yujie Chen, Jun Xue

Abstract: With the vigorous development of computer vision, oriented object detection has gradually been featured. However, angle boundary discontinuity and its knowledge distillation have been the bottleneck for rotating detection distillation design. In this paper, a novel knowledge distillation method named Phase-shift encoded knowledge distillation (PseKD) is proposed to improve the accuracy of predicting object orientation. Specifically, we design a phase-shift encoded module (PSEM) to solve the problem of angle periodicity due to rotational symmetry in oriented object detection by mapping angles to phases of different frequencies. Secondly, an angle knowledge distillation strategy (AKDS) is designed to guide lightweight models to learn angle knowledge from high-performance models. Experiments on public datasets demonstrate that the proposed method can effectively solve the various periodic fuzzy problems caused by rotational symmetry in remote sensing oriented object detection.