The paper introduces a novel loss function called Focal Loss, which is designed to address class imbalance in one-stage object detection. Class imbalance is a significant challenge in training one-stage detectors, as they process a large number of candidate object locations, most of which are easy negatives. The Focal Loss dynamically scales the cross-entropy loss, reducing the contribution of easy examples and focusing more on hard, misclassified examples. This approach enables the training of highly accurate one-stage detectors, such as RetinaNet, which outperforms both one-stage and two-stage detectors on the COCO dataset. The paper also discusses the design of RetinaNet, including its backbone network (Feature Pyramid Network), anchor boxes, and subnetworks for classification and box regression. Experimental results show that the Focal Loss significantly improves the performance of one-stage detectors, achieving state-of-the-art accuracy while maintaining or improving speed over existing methods.The paper introduces a novel loss function called Focal Loss, which is designed to address class imbalance in one-stage object detection. Class imbalance is a significant challenge in training one-stage detectors, as they process a large number of candidate object locations, most of which are easy negatives. The Focal Loss dynamically scales the cross-entropy loss, reducing the contribution of easy examples and focusing more on hard, misclassified examples. This approach enables the training of highly accurate one-stage detectors, such as RetinaNet, which outperforms both one-stage and two-stage detectors on the COCO dataset. The paper also discusses the design of RetinaNet, including its backbone network (Feature Pyramid Network), anchor boxes, and subnetworks for classification and box regression. Experimental results show that the Focal Loss significantly improves the performance of one-stage detectors, achieving state-of-the-art accuracy while maintaining or improving speed over existing methods.