• We proposed the Local Training (LT) strategy, which is an auxiliary training-only stream that focus training gradients on tiny targets. Consistent improvements across all YOLO variants, without modifying the model architecture or increasing inference cost, making it easy to integrate into real-time systems. • P–Ł convergence analysis and gradient decomposition show that LT boosts target-focused gradient signals and accelerates training convergence, providing practical improvements in detection efficiency. YOLO-based object detection algorithms have consistently advanced one-stage vision in real-time systems. However, for tiny object ( ≤ 16 × 16 pixels) detection performance gaps remain critical in engineering scenarios such as surveillance and anti-UAV detection, where limited computational resources further restrict accuracy and stability. In this paper, we revisit the training dynamics of several representative YOLO versions and uncover key limitations when handling extremely small targets. Building on this, to meet engineering requirements, we proposed a simple yet effective dual-path training strategy named Local Training. Without altering the network architecture, our approach introduced a label-guided auxiliary stream that injects semantic priors into the training input, boosting convergence and enabling the network to better capture small object cues. Theoretically, we show that the dual-loss optimization satisfies the Polyak-Łojasiewicz condition, supporting the fast convergence of YOLO series. Experiments are conducted across YOLOv8 to the latest YOLOv13 on both multi-scale object dataset SOD4bird and tiny object dataset Airborne object tracking (AOT) to validate the efficiency of our method. On SOD4bird dataset, the proposed strategy accelerates the training process and improves mean Average Precision at IoU threshold 0.5 (mAP@50) and recall on the best YOLO version. On AOT benchmark, our approach achieves notable performance with consistent gains in recall and mAP@50. For more challenging TOD scenarios (objects smaller than 16 × 16 pixels), our method attains SOTA performance, yielding 50.5% and 58.0% relative improvements in recall and mAP@50 for YOLOv8n, respectively. Importantly, these results introduce no extra latency or computational burden, showing that LT provides a general, plug-and-play training paradigm that enhances tiny object detection while remaining deployment-friendly for real-time engineering applications.
Gui et al. (Sun,) studied this question.