They're implemented this particular way to provide training data for image segmentation systems, they move the image around inside the frame which allows them to use a few people doing the challenge to create a boundary representation that can be used to train things like YOLO style ML systems
