Learnable representations that focus on individual objects within a scene, capturing their properties and relationships separately.