Advancing Object Detection in Remote Sensing: a Rigorous Evaluation of YOLO Models

Detecting or locating objects in high-resolution aerial imagery is a difficult process owing to the occurrence of small, rotated, and multi-scale objects. The YOLO series of deep neural networks has been the most frequent choice for real-time object detection for remote sensing. The four YOLO models, YOLOv4, YOLOv5, YOLOv7, and YOLOX, are compared on the DOTA dataset in this paper to determine their efficiency in terms of detection of objects with various orientations and scales. Comparison is based on the performance parameters of inference speed, mAP, and computational cost. The conclusion is that the YOLOv7 performs the better Accuracy-Speed tradeoff while the YOLOX performs better for the detection of rotated objects because of the implementation of an anchor-free mechanism. The research work in this paper will help in making the right decision for the YOLO model in real applications of remote sensing.