Overview of single-stage object detection models: from Yolov1 to Yolov7

Abstract

Recently, with the rise of Artificial Intelligence (AI) and computer vision techniques, object detection using visual automated methods has attracted much attention from computer vision researchers. It is leveraged and integrated in a wide range of real-time applications in view of the reliability of the delivered information. The visual detection relies on processing the appearance and motion features received from Electro-optical (EO) and thermal sensors. Moreover, the detection methods have known a significant evolution in terms of methods used, they moved from using handcrafted feature-based methods to deep learning algorithms with two-shot and single shot object detectors. Real time applications use Single Shot object Detectors (SSD) thanks to the optimal compromise that they present between speed and performance in comparison to their counterpart. You Only Look Once (YOLO) algorithms have demonstrated high performance and fast inference speed in several applications, thus they are extensively exploited and improved properly from the first version. This paper present an exhaustive overview and comparison of Yolo algorithms including the major contributions of each version as well as their evolution with respect parameters used and the followed architecture.