The metrics of custom dataset is abnormal

[MrPanda007]

I trained the model using my own dataset, with the detection categories being person, car, and bike, and the training set size is 1 million samples. However, compared to YOLOv5, the training results show that for small-sized targets (between 10-20 pixels), the recall rate of person has increased by 16 percentage points, while those of car and bike have both decreased by 8 percentage points each. What might be the cause of this, and are there any optimization suggestions?