How do you split the dataset into train/test making sure each classes have the same distribution of the classes for train/test split since the random sampling won't suffice this? Especially if you're dealing with multi class object detection.