Figure 1 for Unsupervised Meta-learning of Figure-Ground Segmentation via Imitating Visual Effects