representation for 3D object detection and the template-based prediction, as well as the significantly improved