University of Amsterdam

Contrastively-trained Structured World Models (C-SWMs) depart from traditional pixel-based reconstruction losses and use an energy-based hinge loss for learning object-centric world models.