#retinanet

During training, one image yields 100,000 easy background anchors, each classified correctly with probability $p_t = 0.99$, and 100 hard anchors with $p_t = 0.3$. Under the cross-entropy loss $-\log p_t$, the easy anchors contribute a fraction of the total loss. Under the focal loss $-(1 - p_t)^2 \log p_t$, they contribute a fraction .