Yolov1论文理论部分:https://blog.csdn.net/u011974639/article/details/78208773
Yolov2论文理论部分:https://blog.csdn.net/u011974639/article/details/78208896
https://blog.csdn.net/litt1e/article/details/88852745
K-means聚类:(使得anchor box无限靠近bbox)
https://blog.csdn.net/xiaomifanhxx/article/details/81215051?ops_request_misc=&request_id=&biz_id=102&utm_term=yolov2%20kmeans&utm_medium=distribute.pc_search_result.none-task-blog-2allsobaiduweb~default-1-81215051
Yolov3论文理论部分:https://blog.csdn.net/litt1e/article/details/88907542
FPN解析:(提高小目标检测精度)https://blog.csdn.net/u011974639/article/details/78244743
Tips:
①这里注意bounding box 与anchor box的区别:
Bounding box它输出的是框的位置(中心坐标与宽高), confidence以及N个类别。anchor box只是一个尺度即只有宽高。
②softmax用来处理分类问题,logistic回归问题用来处理是和不是的问题。yolov3将logistic替换原先的softmax,就是为了排除例如检测出男人也属于人这一类,遍历所有类别(是或不是),选取大于阈值的即可。
https://blog.csdn.net/weixin_43384257/article/details/100974776?ops_request_misc=%257B%2522request%255Fid%2522%253A%2522158977991219195265912634%2522%252C%2522scm%2522%253A%252220140713.130102334.pc%255Fall.57693%2522%257D&request_id=158977991219195265912634&biz_id=0&utm_medium=distribute.pc_search_result.none-task-blog-2allfirst_rank_v2~rank_v25-1-100974776.nonecase&utm_term=yolo%E7%BD%AE%E4%BF%A1%E5%BA%A6confidence