Self-supervised Multi-view Disentanglement for Expansion of Visual Collections