The focal-loss curve says what one example costs; this says what the whole image costs. The population on the left is an illustration with its distribution stated, not a measurement: a hundred thousand anchors, almost all of them background the model already gets right at about p_t = 0.9. Before you move anything: at γ = 0 and the published α = 0.25 the twelve positives pay 0.046% of the total loss. Push γ up and watch which block loses the budget — and notice that it is the easy background that gives way, not the background as a whole. RetinaNet divides this total by the number of positive anchors; that scales the bar and changes no share.