Focal Loss for Dense Object Detection

Focal loss discounts already-correct predictions so many easy background candidates do not overwhelm a dense detector. Calculate the loss contribution, separate class weights from difficulty weights, and inspect RetinaNet’s controlled comparison.

Reading focus: Why many small easy-background losses can dominate a dense detector’s training objective. How the correct-class probability controls focal weighting and how gamma differs from alpha. What the matched RetinaNet ablation establishes and why hard examples can still contain bad labels.

ICCV 2017. Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He, and Piotr Dollár. 12 min read, easy difficulty.