For this demonstration, the network's dog probabilities are averaged over 32 × 32 blocks of the photograph, one value per block, the dots on the left. Every pixel needs a label, so something has to fill in between them. Before switching: which will follow the dog's outline better, enlarging the coarse labels or interpolating the scores first, and will either recover the legs? The scores are real output of torchvision's FCN-ResNet50, which itself predicts at stride 8; the outline is Mask R-CNN's dog mask, used as a reference rather than as ground truth.