Feature Adaption with Predicted Boxes for Oriented Object Detection in Aerial Images

Minhao Zou, Ziye Hu, Yuxiang Guan, Zhongxue Gan, Chun Guan, Siyang Leng

Abstract: Object detection is a fundamental research field in computer vision. Arbitrary-oriented objects inevitably appear in face, natural scene text, and aerial image detection, which have attracted widespread attention recently. However, existing rotation detectors still suffer from the feature misalignment problem, due to the fixed convolution kernel adopted in detecting arbitrary-oriented and deformed objects. In this paper, we propose a novel method, One-stage Feature Adaption Network (OFA-Net), for oriented object detection in aerial images. A feature adaption module, implemented by the deformable convolution and the align convolution, is proposed to refine the feature maps according to the predicted offsets and decoded boxes. Furthermore, specific to the long-existing periodic angle regression problem in the detection, the box regression branch is decoupled into the size branch and the angle branch, with a new periodic loss in the angle regression branch to leverage the periodic orientation of the object. Extensive experiments demonstrate the effectiveness of our approach, achieving promising results compared with state-of-the-art methods in three benchmark datasets, DOTA, HRSC2016, and UCAS-AOD.