An FCN's last layer is a 1 × 1 convolution with one output channel per class, so its output is a volume: one score map per class, stacked in depth. Move over the volume, or down the list, to pull a channel out and see where its class wins. The illustration is built by hand, with six classes you can name, to show the idea; it is not network output. Then switch to the real scores of torchvision's FCN-ResNet50, trained on the 21 Pascal VOC classes: before you do, predict what it can say about the sky, the tree and the grass, which are not among them.