is the model retinaface trained on RGB/BGR way?

xinntao / facexlib

FaceXlib aims at providing ready-to-use face-related functions based on current STOA open-source methods.

MIT License

810 stars 143 forks source link

is the model retinaface trained on RGB/BGR way? #15

Open asdf1996 opened 2 years ago

asdf1996 commented 2 years ago

Hi, i'd like to know is the model retinaface trained on RGB or BGR? if i applied direct on RGB, will the precision changes a lot? Thanks for reply!

woctezuma commented 2 years ago

https://github.com/ternaus/retinaface/issues/27

asdf1996 commented 2 years ago

@woctezuma Thanks for quick reply! so in this repo,the model is trained on RGB without changing the network structure? Can I apply the pretrained weights in this repo to facexlib?

woctezuma commented 2 years ago

It has been a long time, but I would say that, based on the issue linked above:

biubug6/Pytorch_Retinaface uses BGR,
ternaus/retinaface uses RGB.

Based on the README of xinntao/facexlib, the model comes from biubug6/Pytorch_Retinaface, so it should be BGR as well.

I believe that it would be better if:

you did not need to import the weights from ternaus/retinaface,
you could simply stick to the BGR convention when using xinntao/facexlib.