hambox: Delving into Online High-quality Anchors Mining for Detecting Outer Faces
done :
1 online high quality achor mining a little different from the paper,assume there is an outer face ,compansented anchor and anchor matched in the fisrt step should be top K in candidates, the sort algorithm may be difficult，so after calculte iou outerface and decoded regression,erase M anchors which match in the first step,the sort the left anchors and select k- m as compensated anchors
2 regression focal loss
3 data anchor sample
todo
1 paper refers use p2-p6 five layers,anchor size is 0.68*[16,32,64,128,256,512], re-implement anchor size dont take ratio 0.68 into account
2 train loss is convergence , but during training ,dead lock always appears,the training code is under going
change date 1103
many useless code and some error，the compensated anchor score should with no grad， refer to c_iou
base project https://github.com/biubug6/Pytorch_Retinaface/ just use hamloss insteat of multiboxloss