search

YongtaoGe / RetinaFace

Reimplement RetinaFace using PyTorch.
MIT License
74 stars 15 forks source link

a litte problem in fpn code #4

Open foocker opened 5 years ago

foocker commented 5 years ago

in line 45: p3 = F.interpolate(p4, size=[p4.size(2), p4.size(3)]) + p4 should change to: p4 = F.interpolate(p5, size=[p4.size(2), p4.size(3)]) + p4

foocker commented 5 years ago

there is no Context Module, so has no DCN operator in your RetinaFace Module. i have a problem: when training, all prediction bbox and landmark was catted in one dim, how to get the scale level of them for loss function's forward. thx.

YongtaoGe commented 5 years ago
  1. yes, you are right!
  2. I just couldn't understand your second question.
foocker commented 5 years ago

maybe something that i am not understand well, i mean like the two-stage, all predicted anchors should be get its own level from the FPN after gathered them together, only this we can write the loss function for backward. it seems Retina + FPN have no such worry.

yongtaoge notifications@github.com 于2019年8月18日周日 下午7:12写道：

  1. yes, you are right!
  2. I just couldn't understand your second question.

— You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub https://github.com/yongtaoge/RetinaFace.PyTorch/issues/4?email_source=notifications&email_token=AFHSUMJWUZJAVVURHTSPNF3QFEVBFA5CNFSM4IMRIAW2YY3PNVWWK3TUL52HS4DFVREXG43VMVBW63LNMVXHJKTDN5WW2ZLOORPWSZGOD4Q5VLI#issuecomment-522312365, or mute the thread https://github.com/notifications/unsubscribe-auth/AFHSUMNDIH4R76OUJDYKKDDQFEVBFANCNFSM4IMRIAWQ .