Unsupervised learning of object structure and dynamics from videos