What is Focal Loss and why is it important in object detection?

Focal Loss reduces the impact of easy negatives and focuses training on hard examples in imbalanced detection tasks.

In object detection, background examples dominate foreground objects, creating extreme class imbalance. Focal Loss modifies cross-entropy by adding a modulating factor (1 - p)^γ, which down-weights easy examples and focuses learning on hard misclassified samples. This is crucial in dense detectors like RetinaNet, where negative anchors far outnumber positive ones.