Rotate-Yolov5 for Aerial Images

Abstract

In Recent years, great progress has made in object detection. However, since the orientation of object in aerial image is random, the regular horizontal object detection method is not suitable for aerial images. In this paper, we present a Rotate-Yolov5 network based on Yolov5. We use an Adaptive Rotating Anchor Generation Module (ARAGM) to generate anchors with object orientation information. Then the orientation information is used for Rotate-Deformable Convolution Module (R-DCM) to extract features. Finally, we use a decouple detection head as Oriented Object Detection Module (OODM) to yield classification and regression results. Moreover, Rotate-Smooth L1 is used to optimize the loss function. We evaluate the proposed Rotate-Yolov5 on DOTA datasets and the mAP reached 75.4, which demonstrate the superiority of its effectiveness.