Contrastive Training of Complex-Valued Autoencoders for Object Discovery