Synchronized Batch Norm跨GPU批归一化策略最早在[MegDet: A Large Mini-Batch Object Detector](https://arxiv.org/abs/1711.07240)
论文中提出,在[Bag of Freebies for Training Object Detection Neural Networks](https://arxiv.org/pdf/1902.04103.pdf)论文中以Yolov3验证了这一策略的有效性,[PaddleCV/yolov3](https://github.com/PaddlePaddle/models/tree/develop/PaddleCV/yolov3)实现了这一系列策略并比Darknet框架版本在COCO17数据上mAP高5.9.
论文中提出,在[Bag of Freebies for Training Object Detection Neural Networks](https://arxiv.org/pdf/1902.04103.pdf)论文中以Yolov3验证了这一策略的有效性。