| Title |
Evaluating YOLOv11 for Real-time Small Object Detection: Accuracy?efficiency Pareto Analysis |
| DOI |
https://doi.org/10.5573/ieie.2025.62.12.150 |
| Keywords |
YOLOv11; Small object detection; GFLOPs; mAP@0.5; Pareto frontier |
| Abstract |
This study systematically evaluates the performance of YOLOv11n for small object detection. The proposed model comprises 2.6M parameters and 6.50 GFLOPs at an input size of 640×640, demonstrating computational efficiency suitable for real-time applications. Experimental results indicate that YOLOv11n achieves the highest accuracy with mAP@0.5=18.76%, while YOLOv5n demonstrates superior resource efficiency with a model size of 3.87 MB and 156 FPS. The two models occupy different points on the Pareto frontier, highlighting their complementary applicability for accuracy-oriented and efficiency-oriented scenarios. Ablation studies confirm the contributions of the C3k2 block, DFL loss, and Anchor-free regression to performance improvement. Robustness experiments further show that YOLOv11n maintains R ≥ 85% across diverse conditions, ensuring stable detection. This work contributes by delivering a comprehensive quantitative assessment and deployment guidelines for small object detection, while future research will explore ultra-small object enhancement, on-device optimization, and domain adaptation for real-world applications. |