we define H, which maps the predicted distribution Z to point estimate Y in ab space. One choice is to take the mode of the predicted distribution for each pixel, as shown for in the right column of Figure 4 for some qualitative examples. This provides a vibrant but sometimes spatially inconsistent result. For example, pieces of the tent in the third image are red, cyan, and orange, and the fourth image presents a salient blue artifact in the top left corner. On the other hand, taking the mean or expectation of the predicted distribution, shown in the left column, results in a spatially consistent but desaturated prediction. For example, the table and the purple flowers in the first two images take on an unnatural sepia tone. This is unsurprising, as taking the mean after performing classification suffers from some of the problems as optimizing for a Euclidean