Rotated-DETR: an End-to-End Transformer-based Oriented Object Detector for Aerial Images
G.S. Lee, Jinbeom Kim, Taejune Kim, Simon S. Woo
Abstract
Oriented object detection in aerial images is a challenging task due to the highly complex backgrounds and objects with arbitrary oriented and usually densely arranged. Existing oriented object detection methods adopt CNN-based methods, and they can be divided into three types: two-stage, one-stage, and anchor-free methods. All of them require non-maximum suppression (NMS) to eliminate the duplicated predictions. Recently, object detectors based on the transformer remove hand-designed components by directly solving set prediction problems via performing bipartite matching, and achieve state-of-the-art performances in general object detection. Motivated by this research, we propose a transformer-based oriented object detector named Rotated DETR with oriented bounding boxes (OBBs) labeling. We embed the scoring network to reduce the tokens corresponding to the background. In addition, we apply a proposal generator and iterative proposal refinement module in order to provide proposals with angle information to the transformer decoder. Rotated DETR achieves state-of-the-art performance on the single-stage and anchor-free oriented object detectors on DOTA, UCAS-AOD, and DIOR-R datasets with only 10% feature tokens. In the experiment, we show the effectiveness of the scoring network and iterative proposal refinement module.