A one-stage detector emits one tensor, and every cell of it means something. Click a cell of the grid to see its slice written out. The cell holding the dog's centre is the one YOLO makes responsible for the dog, computed from the hand-drawn ground-truth box; the numbers under the slots are the training targets that box implies, not a network's output, so nothing here is invented. Change B and watch the tensor grow while the picture does not. Then switch to the anchor form: the same boxes, but the class scores stop being shared, so the tensor gets larger for the same number of predictions. The photograph's other object, the ball, is left out — sports ball is not one of VOC's twenty classes.