Self-supervising amodal 3D scene feature representations