Why not add the class of background in training?

[XRDai]

For object detection in images, the background is added as a class, such as in COCO object detection, there are 80 classes, but the output class is 81(+backgroud). However, in SECOND, there are only three classes are predicted in KITTI (Car, Ped, Cyc), why backgroud is not added for a class in prediction?