Minimalist notes: Focal Loss

"Minimalist Notes" Focal Loss

Focal Loss for dense Object detection

The core of the article is to propose the focal loss to replace the original cross entropy loss
Pt={p,1−p,if y=1otherwise p t = {p, if y = 1 1−p, otherwise p_t=\begin{cases}p, & \text{if $y =1$} \ \ 1-p, & \text{otherwise}\end{cases}
CE (PT) =−log (PT) C E (P t) =−log (P t) ce (p_t) =-\log (p_t)
FL (PT) =− (1−PT) Γlog (PT) F L (P t) =− (1−p t) γlog (P T) FL (p_t) =-(1-p_t) ^\gamma\log (p_t)

The Γγ\gamma in focal loss is an adjustable parameter, and the γ=2γ= 2 \gamma=2 is the best in the experiment. The authors also try to add Αα\alpha Balance effect loss function FL (PT) =−α (1−PT) Γlog (PT) F L (P t) =−α (1−p t) γlog (P T) FL (p_t) = -\alpha (1-p_t) ^\gamma\log (p_t), when