focal loss

**Focal loss** is a **modified cross-entropy loss function designed to address extreme class imbalance** — by down-weighting well-classified (easy) examples and focusing training on hard, misclassified samples, focal loss enables single-stage object detectors like RetinaNet to achieve accuracy comparable to two-stage detectors. **Why Focal Loss Matters** - **Class Imbalance**: In object detection, background patches outnumber objects 1000:1. - **Easy Example Problem**: Standard cross-entropy wastes gradient on trivially classified negatives. - **Hard Mining Alternative**: Focal loss automates what hard negative mining does manually. - **Single-Stage Detectors**: Made RetinaNet competitive with Faster R-CNN. **Formula** FL(p) = −α(1 − p)^γ log(p), where γ (gamma, typically 2) controls the focusing strength and α balances class weights. **Impact by Example Difficulty** - Easy example (p=0.9): Loss reduced to 0.01% of standard CE. - Hard example (p=0.1): Loss remains at ~90% of standard CE. - Result: Model focuses almost entirely on difficult, informative examples. **Applications**: YOLO, RetinaNet, SSD, medical image segmentation, fraud detection, any task with severe class imbalance. Focal loss **transformed object detection** — proving that class imbalance, not architecture, was the main barrier to single-stage detector performance.