AFPN Based YOLOX for Rotation Object Detection in Remote Sensing Image

Rotation Object Detection in Remote Sensing Image is a big challenge, as it has different sizes and orientations. In this paper, we proposed an AFPN and YOLOX based Rotation Object Detection method, named AFPN-RYOLOX. In the AFPN-RYOLOX, an angle prediction branch added into the detection head of YOLOX and revised the loss function to incorporate Kullback-Leibler Divergence (KLD) loss firstly, which enhanced loss continuity and ensuring scale invariance. Then, the Asymptotic Feature Pyramid Network (AFPN) is integrated in the neck layer of YOLOX, which boosts the network's capability to extract features and decreases the number of model parameters through a dual-phase adaptive spatial fusion process. Through compare with other SOTA method, the proposed AFPN-RYOLOX achieved superior performances on DOTA and HRSC2016 datasets.