Page 1 of 1

Q - extracting *only* a training set from swap b

Posted: Sun May 21, 2023 1:22 pm
by ben

i followed the guide several times.
just want to know if i understand.
when extracting only a training set from swap b, i just extract every N - half a sec to a sec frames, with the desirable mask? and thats it?

considering there is only 1 face in swap b and no need to sort and stuff, is there no need to firstly fix alighnments for swap b? with the manual tool and so on? and then extract the swap b's faces from the fixed alighnment file?

is the mask applied correctly on extracted faces with no any alighnments?

i know that there are a lot of swapping options but i gave this example just for the sake of this question

thanks


Re: Q - extracting *only* a training set from swap b

Posted: Mon May 22, 2023 6:40 pm
by torzdf
ben wrote: Sun May 21, 2023 1:22 pm

when extracting only a training set from swap b, i just extract every N - half a sec to a sec frames, with the desirable mask? and thats it?

Yes.

considering there is only 1 face in swap b and no need to sort and stuff, is there no need to firstly fix alighnments for swap b? with the manual tool and so on? and then extract the swap b's faces from the fixed alighnment file?

Yeah, I generally don't bother fixing up just training data. I just visually scan through the extracted faces and delete any obviously misaligns/false-positives

is the mask applied correctly on extracted faces with no any alighnments?

If you select the required mask at extract time, this will be fine, yes. Otherwise you can add masks on with the Mask tool.


Re: Q - extracting *only* a training set from swap b

Posted: Tue May 23, 2023 8:11 am
by ben

is the mask applied correctly on extracted faces with no any alighnments?

If you select the required mask at extract time, this will be fine, yes. Otherwise you can add masks on with the Mask tool.

from the guide: "Mask Generation - Identify the parts of an aligned face"
so as said, the mask is generated according to the alignment. so if i dont fix alignments it might end with masks not aligned correctly on the face..?

you scan it with the manual tool or in the folder?

but from what you said for training i can just extract straight from the source with the mask and not from an alignment file..
just trying to understand

thanks


Re: Q - extracting *only* a training set from swap b

Posted: Thu May 25, 2023 11:33 am
by torzdf
ben wrote: Tue May 23, 2023 8:11 am

is the mask applied correctly on extracted faces with no any alighnments?

If you select the required mask at extract time, this will be fine, yes. Otherwise you can add masks on with the Mask tool.

so as said, the mask is generated according to the alignment. so if i dont fix alignments it might end with masks not aligned correctly on the face..?

If you use landmarks based masks (extended/components) then it is generated from the alignments, yes, in which case you would want the landmarks to be correct. If you use NN based masks (eg. bisenet-fp) then the mask is generated directly from the aligned face image. The landmarks do not enter in to it at all.


Re: Q - extracting *only* a training set from swap b

Posted: Sat May 27, 2023 7:19 pm
by ben

ookk

thanks!


Re: Q - extracting *only* a training set from swap b

Posted: Tue Jun 27, 2023 10:05 am
by torzdf

I'm not sure what you mean by more complicated? It just generates the masks a different way (by filling in the area between landmark points, rather than using a Neural Network to create the mask). All the masks are stored exactly the same way, regardless of how they were originally created