Anfonwch hwn fel neges destun: Self-supervised learning of structural representations of visual objects