Abstract
Object detection in remote sensing consists in recognizing and localizing objects of interest on the Earth’s surface. Classical object detectors must deal with a wide range of distinct side-views, generally oriented upward due to gravity, while objects in aerial images appear in a quasi-standard bird-eye view with an arbitrary orientation and low resolution on the ground plane. This naturally leads to the question: Can we detect the direction of an object along with its location and object category? So-called oriented bounding boxes (OBB) became standard in remote sensing due to the crowded scenes, especially for rotated and elongated objects (e.g., parking cars in a parking lot). OBBs localizes objects by a minimal enclosing box, thus oriented refers to a rotation angle of a standard BB for a tight enclosing of the detected object with respect to the image axes. However, this bounding box orientation does not correspond to the orientation of the detected object, e.g., the front of an airplane or a car, the stem of a leaf, which provides valuable information for aerial image analysis. Herein, we propose a novel method, the Directed Object Detector (DOD), which is capable of detecting the object’s direction together with its minimal enclosing OBB. This is integrated into an object detector to create a single end-to-end neural network. Experimental validation confirms the state of the art performance on both close and far-range remote sensing images in man-made and natural environments. Furthermore, we prove the advantage of DOD over the OBB approach in an image rectification application, which is not solvable with OBB.
