RRNet: A Hybrid Detector for Object Detection in Drone-Captured Images

Objects captured by UAVs and drones in city scenes usually come in various sizes and are extremely dense. Therefore, we propose a hybrid detector, called RRNet, for object detection in such challenging tasks. We mix up the anchor-free detectors with a re-regression module to construct the detector. The discard of prior anchors released our model from the difficult task on bounding-box size regression so that we achieved a better performance in multi-scale object detection in the dense scene. The anchor-free based detector firstly generates the coarse boxes. A re-regression module is then applied on the coarse predictions to produce accurate bounding boxes. In addition, we introduce an adaptive resampling augmentation strategy to logically augment the data. Our experiments demonstrate that RRNet significantly outperforms all the state-of-the-art detectors on VisDrone2018 dataset. We are runner-up to the ICCV VisDrone2019 Object Detection in Images Challenge, and we achieve the best AP50, AR10, and AR100. Source code will be published on our official website in due course.