Mosaic data augmentation technique introduced in YOLOv4 paper implemented in Pytorch. In this method four images are combined into a grid and then randomly cropped. The main idea is to teach a model to identify objects outside of their usual environment.
Find more here.