Figure 3. Illustration of localization distillation (LD) for an edge e ∈ B = {t, b, l, r}. Only the localization branch is visualized here.
S (·, τ ) is the generalized SoftMax function with temperature τ . For a given detector, we first switch the bounding box representation
to probability distribution. Then, we determine where to distill via region weighting on the main distillation region and the valuable
localization region. Finally, we calculate the LD loss between two probability distributions predicted by the teacher and the student.