Seol mar théacs é seo: Self-supervised learning of structural representations of visual objects