Building instance detection models that are data efficient and can handle rare object categories is an important challenge in computer vision. But data collection methods and metrics are lack of research towards real scenarios application using neural network. Here, we perform a systematic study of the Object Occlusion data collection and augmentation methods where we imitate object occlusion relationship in target scenarios. However, we find that the simple mechanism of object occlusion is good enough and can provide acceptable accuracy in real scenarios adding new category. We illustate that only adding 15 images of new category in a half million training dataset with hundreds categories, can give this new category 95% accuracy in unseen test dataset including thousands of images of this category.