机器学习36:YOLOV4相关理论知识整理:
YOLOV4论文:https://blog.csdn.net/Sophia_11/article/details/105726907
YOLOV4论文翻译:https://blog.csdn.net/qq_34795071/article/details/105734300?fps=1&locationNum=2
机器之心总结:https://www.jiqizhixin.com/articles/2020-04-24-4
极市平台总结:https://blog.csdn.net/Extremevision/article/details/105731159
1.图示:
2:最终方案:
(1)CSP Darknet53:
检测模型需要具有以下特性:
1)更高的输入分辨率,为了更好的检测小目标;
2)更多的层,为了具有更大的感受野;
3)更多的参数,更大的模型可以同时检测不同大小的目标。
也就是选择具有更大感受野、更大参数的模型作为backbone。下图可以看到:CSPResNeXt50仅仅包含16个卷积层,其感受野为425x425,包含20.6M参数;而CSPDarkNet53包含29个卷积层,725x725的感受野,27.6M参数。这从理论与实验角度表明:CSPDarkNet53更适合作为检测模型的Backbone。
(2)SPP:
https://www.cnblogs.com/zongfa/p/9076311.html
(3)PANet:
https://www.sohu.com/a/225356135_129720
(4)YOLOv3 Head
(5)CmBN:创新点
(6)多输入加权残差连接(Multi-input weighted residual connections):
在bifpn中,提出了多输入加权残差连接以执行按比例的级别重新加权,然后添加不同比例的特征图。
bifpn论文: https://arxiv.org/pdf/1911.09070.pdf
(7)Mish:
https://blog.csdn.net/u011984148/article/details/101444274
(8/9)ciouLoss/diouloss:
https://blog.csdn.net/qq_33270279/article/details/103482085
3.训练方法:数据增强:
(1)Mosaic:创新点
(2)cutmix:
(3)SAT:创新点
(4)Label Smoothing:
https://blog.csdn.net/qq_43211132/article/details/100510113
(5)dropblock:
https://blog.csdn.net/qq_14845119/article/details/85103503
4.训练方法:策略:
(1)Cosine annealing scheduler:
https://blog.csdn.net/CarryLvan/article/details/104394960
https://blog.csdn.net/zisuina_2/article/details/103250274
(2)Random Training Shapes:
https://blog.csdn.net/qq_33869371/article/details/103076561
https://www.cnblogs.com/Terrypython/p/10642091.html