x-axis: the probability the model gives the correct class (pt). Cross-entropy (γ = 0) still gives easy, well-classified examples (pt > 0.6) a noticeable loss; focal loss multiplies it by (1 − pt)γ, which shrinks it toward 0.
10 detections of one class, sorted by confidence, for an image set with 6 real objects. Click a detection to mark it correct (TP: IoU ≥ 0.5 with an unmatched object) or wrong (FP).