You have trained a custom object detection model using Azure AI Vision and evaluated its performance using the Mean Average Precision (mAP) metric. You find that the mAP is low for some classes of objects. What could be the possible reasons for this?