search

biubug6 / Pytorch_Retinaface

Retinaface get 80.99% in widerface hard val using mobilenet0.25.
MIT License
2.63k stars 774 forks source link

hard gt #112

Open mike112223 opened 4 years ago

mike112223 commented 4 years ago

Hi, I noticed that you labeled hard gts as 2 to ignore, but when computing loss those anchors attached to hard gt still be considered. So actually anchors are either positive or negative. Am i right? I am afraid that I missed some important info. Thank you!!!