One-stage oriented object detection in remote sensing images

Advances in technology resulted in an enormous amount of information collected from high technology satellites and aircraft sensors. These high-resolution images obtained by the said platforms enabled humans to understand the Earth better. The first studies in the field of remote sensing image analysis have focused on image classification, which studies the understanding of scenes, terrain, and vast fields. Recently object detection and image segmentation on very highresolution electro-optical images have become popular. Modern object detectors have two types of architecture. Two-stage object detectors screen probable object locations with a region proposal mechanism and regress over these preselected regions to classify objects. On the other hand, onestage object detectors perform global regression and classification over the image without any preselection. The region proposal mechanism improves the accuracy but slows the detection speed, making one-stage object detectors better for realtime detection, which have lower accuracy. Publicly available data sets generally provide horizontal bounding box labels. However, remote sensing images are dominated by small, congested, or large aspect ratio objects, which is suitable for oriented bounding boxes. We study one-stage object detection in one of the most extensive remote sensing object detection data sets (DOTA), which includes fifteen object classes. Our one-stage object detectors are based on one of the most prominent one-stage object detector architectures (YOLO). Horizontal and oriented bounding box labels are used to obtain results to show that the oriented object detection setting has a surpassing performance over axis-aligned object detection setting in general and also to show that one-stage object detection can have competitive results compared to two-stage object detection methods. Our oriented object detection approach converts angle regression into a form of classification and is easily applicable to the one-stage detectors. Our experiments show that rotated object detection based on angular classification performs better than horizontal object detector. It also shows 3.07% improvement over published two-stage methods and 12.84% improvement over published anchor-free methods on the validation data set.