YOLO11m Precision and Recall Plateau Around 0.50 After 50 Epochs

I am training a YOLO11m object detection model for Pakistani road-traffic vehicle detection using a dataset of approximately 1,000–5,000 images and 21 classes. My training configuration is epochs=100 , batch=-1 , and imgsz=640 , using the dataset haris-habib/datasets/road-vehicles . After 100 epochs, I obtained precision of 0.741, recall of 0.485, mAP50 of 0.517, and mAP50-95 of 0.308. My training losses were box loss 0.603, classification loss 0.325, and DFL loss 0.837. My validation losses were box loss 1.321, classification loss 1.231, and DFL loss 1.267. Recall increased from approximately 0.30 to 0.50 during the first 50 epochs, then plateaued; mAP50 also stopped improving. My dataset is imbalanced, with 5,472 car annotations but only 3 garbage-van annotations, alongside several overlapping vehicle categories. This is my third model-training experiment, and I am testing different image sizes, but precision and recall remain unsatisfactory after 100 epochs. What could be causing this plateau, and how can I determine whether the main issue is overfitting, class imbalance, annotation quality, or image resolution? Should I modify the dataset, merge similar classes, adjust training parameters, increase image size, or use a different model architecture?