subtopic

Baseline, then train YOLO

First run a pretrained YOLO (trained on COCO's 80 classes) on your images — it finds 'person' but knows nothing about helmets, which shows why we train. Then fine-tune from pretrained weights for 50–100 epochs. Read the training plots (box loss, class loss, precision, recall, mAP) and the sample predictions YOLO saves.