yolo series of articles principle

https://zhuanlan.zhihu.com/p/32525231 (yolo v1)

https://zhuanlan.zhihu.com/p/35325884 (yolo v2, compared to v1, with many small tricks, including BN; anchors; kmeans-prior clustering; confidence becomes a not a per Prior a grid cell; use of passthrough layer; (tx, ty, tw, th) is calculated constrained; darknet is introduced; multiscale training;)

https://zhuanlan.zhihu.com/p/49556105 (yolo v3, for a mainstream network, using the feature extractor darknet-53, which has a residual internal network, after use the FPN fusion multiscale feature)