DECOUPLED INSTANCES DISTILLATION FOR REMOTE SENSING OBJECT DETECTION

As a significant field in object detection, knowledge distillation is plagued by coarse pixel division and complex object form. Although some methods currently use ground truth boxes for pixel division, they ignore consideration of teacher knowledge, resulting in inferior or even erroneous knowledge not being distinguished. In this paper, we propose a distillation method named Decoupled Instances Distillation, which solves the problem of imprecise pixel division and complex forms of remote sensing objects. First, we use the predicted boxes of the teacher network to distinguish different object instances and discard instances containing wrong knowledge. Second, considering the performance of classification and regression, we design the concentration function to differentiate object instances. Experimental results show that our method is powerful on remote sensing image datasets. Decoupled Instances Distillation improves the performance of YOLOv3 by 3.1% mAP on the DIOR test set, outperforming existing methods.