Optimizers Comparative Analysis on YOLOv8 and YOLOv11 for Small Object Detection

Object detection has seen significant advancements due to deep learning, particularly in applications that require a balance between speed and accuracy. The YOLO (You Only Look Once) series has led developments in real-time object detection, with YOLOv11 offering the latest architectural enhancements for complex environments. While the model architecture is crucial, the choice of optimizer plays a critical role in determining model convergence, precision, and generalization-essential factors for challenging domains such as aerial imagery. In this study, we evaluate several optimizers-AdamW, SGD, Adam, AdaMax, Nadam, and RAdam-regarding their impact on YOLOv11's performance for small object detection in urban settings. Employing the VisDrone dataset, which is renowned for its densely populated urban imagery, we compare the optimizers based on key metrics: mean Average Precision (mAP), precision, recall, and training efficiency. The results reveal that YOLOv11, when paired with AdamW, achieves the highest mAP of 29.6%, precision of 98.1%, and recall of 68.1%, underscoring its exceptional accuracy in object detection. Furthermore, AdamW enables faster convergence, with a training time of 2.42 ms, making it well-suited for real-time UAV surveillance applications. In comparison, SGD demonstrates commendable performance but falls slightly short in precision and recall. This study offers valuable insights into optimizer selection, aimed at enhancing YOLO model performance, especially in demanding tasks involving small object detection.