Comparative analysis of YOLOv9–YOLOv12 object detection models on a custom non-open dataset

Abstract

This paper compares YOLOv9–YOLOv12 object detection models on a custom non-open dataset (6,500 images, 11 classes), all trained for 300 epochs under identical settings. Results show YOLOv12 achieves the highest mAP@0.5 (0.678), outperforming YOLOv9 (0.648), YOLOv10 (0.603), and YOLOv11 (0.649). Architectural optimizations and training stability drive performance disparities, with YOLOv12 exhibiting superior convergence and precision-recall balance, though small-object detection remains a challenge.