Object Detection in Aerial Images with Orientation Awareness

Object detection plays a vital role for aerial images because of the ease of access to images in any geographical region. Object detection on aerial images can automate many tasks from traffic management to surveillance. The major challenge faced during object detection in areal images is that objects are not associated with vertical height whereas objects are usually height dominant in natural images. Moreover, the orientation in areal images is arbitrary with huge variations in the scale of an object. Existing work performs detection with a bounding box aligned with the axis which limits the algorithm to have an oriented or a rotated bounding box which captures a lot of redundant regions and leads to misclassification especially in the case of densely populated regions. Therefore, in this paper, an object detection approach, that can be resilient to the random orientation of the object, is presented. The proposed mechanism will have the exact spatial information of the object and will increase the detection accuracy. The deTection in Aerial Images (DOTA) dataset is employed for evaluating the performance of the proposed algorithm.