Improving YOLO-Based Traffic-Sign Detection on GTSDB via F1-Guided Thresholding and Confusion-Aware Analysis

This paper presents a practical procedure to improve YOLOv10-based traffic-sign detection on the GTSDB benchmark by coupling F1-guided operating-point selection with a confusion-aware error audit. A standard YOLOv10 detector is trained with the reference recipe and evaluated under COCO metrics. Rather than reporting only full precision-recall curves, we derive a deployable confidence threshold from the validation F1-confidence profile and apply it consistently for test evaluation and qualitative analysis. The selected operating point, defined by a tuned confidence threshold of 0.38, attains an F1 score of 0.68. At the set level, the model reaches mAP at IoU 0.50 of 0.92 (best at epoch 105) and mAP averaged over IoU 0.50 to 0.95 of 0.70 (best at epoch 126), with final-epoch values remaining close, indicating stable convergence. Normalized and count-based confusion matrices at this operating point show residual errors dominated by near-neighbor substitutions among adjacent speed-limit denominations and visually similar prohibitions, consistent with the prevalence of tiny objects. The proposed workflow translates benchmark curves into a concrete deployment rule while connecting aggregate scores to actionable diagnostics. Implications include prioritizing resolution-aware inference or cropping for tiny numerals and targeting data curation and augmentation toward the most confusable families. The approach is simple to reproduce and can be adopted alongside existing YOLO pipelines for safety-critical perception.