TARDet: Two-stage Anchor-free Rotating Object Detector in Aerial Images
Longgang Dai, Hongming Chen, Yufeng Li, Caihua Kong, Zhentao Fan, Jiyang Lu, Xiang Chen
Abstract
Detection of rotating object in aerial images is a practical and challenging task. Nowadays, most detectors rely on anchor boxes with different scales, aspect ratios and angles for aerial objects that are usually distributed in arbitrary directions and show huge variations in scale and aspect ratios. However, the detection performance of these detectors is very sensitive to the anchoring hyperparameters. To address this issue, in this paper, we propose a Two-stage Anchor-free Rotating object Detector (TARDet). Our TARDet first aggregates feature pyramid context information by a feature refinement module, and generates rough localization boxes in an anchor-free manner by a directed generation module (DGM) in the first stage, and then refines it to a higher quality localization scheme. Furthermore, we design an alignment convolution module to extract alignment features and introduce RiRoI to adaptively extract rotationally invariant features from isovariant features. Finally, we apply a modified fast R-CNN head to generate the final detection results. Our approach achieves state-of-the-art performance on two popular aerial objects datasets, DOTA and HRSC2016.