드론 객체 탐지를 위한 YOLO 조기 종료 하이퍼 파라미터 성능 평가

초록

In recent object detection research, YOLO (You Only Look Once) models have been widely used for their speed and accuracy, however performance variation can still occur under identical training conditions. This study applies a patience-based early stopping method and conducts 30 repeated training runs of YOLOv3, YOLOv8, and YOLOv11 to analyze performance variability. Using drone nadir imagery, the AP50 achieved with early stopping ranged from 0.879 to 0.977 for YOLOv3, 0.846 to 0.967 for YOLOv8, and 0.801 to 0.928 for YOLOv11, indicating that a single run is insufficient for reliable model comparison. Early stopping also yielded higher average performance than the fixed 100 epoch baseline across all models. The convergence epochs ranged from 124 to 818, and the coefficients of determination between convergence epoch and AP50 were low: 0.20 for YOLOv3, 0.02 for YOLOv8, and 0.04 for YOLOv11. This demonstrates that epoch count alone does not account for performance variability. These findings support dynamically adjusting the stopping point based on validation performance rather than relying on a fixed epoch. Overall, the combination of repeated training and early stopping improves the stability and reliability of performance evaluation for YOLO object detection.

키워드