Multi-Scale Oriented Object Detection With Focus Error Ellipse Loss

The loss function and feature extraction framework are essential parts of the algorithm design and significantly affect the accuracy of oriented object detection in remote sensing images. Though considerable progress has been made, there are still challenges left to be explored, e.g., large variations in scales, arbitrary direction, and dense distribution of the objects, which may have some undesirable effects, such as inaccurate object position regression, high false alarm, and miss rate. To address the above problems, we propose a focus error ellipse (FEE) loss function. This function bolsters the detection accuracy by narrowing the distance between the center points of the labeled and predicted bounding boxes based on the error ellipse. For the network part, we carefully crafted two unit modules: a fine-grained and context-augmented module (FCM) and a semantic information regrouping module (SIRM). The FCM aligns fine-grained information with contextual information to establish dependencies between local and global features, which helps to grasp more holistic characteristics of objects. The SIRM reorganizes the acquired deep semantic features in the channel dimension, enhances the weight of task-beneficial semantic information, and further derives the optimal combination method of feature subsets for object detection. Based on the aforementioned work, we developed an oriented object detection framework, which further improves the detection accuracy of large aspect ratio objects and dense scenes. The experimental results show that the proposed method can produce competitive performance in oriented object detection compared to other state-of-the-art models.