Adding 1adrianb/face-alignment to project about faceswap HOT 10 CLOSED

Yes, I have seen that and it looks promising. Do you want to take a look at it ?

There is also a Gist with a sample:
https://gist.github.com/anonymous/5832a6fffb8309ac851a29874c3dbb7b

from faceswap.

Btw, note that face recognition seems to be better with the face_recognition module integrated here. So the dlib.get_frontal_face_detector() should be replaced where relevant

from faceswap.

I pushed a first draft on my repo, inside the align folder, it is the GIST code reworked to separate program logic from processing logic.

It is not intended to work, but if you want to work on it, please start from there.

CC @Ganonmaster

from faceswap.

@Clor would love to but I can't nowadays. coming months I will be helping other issues and features though

from faceswap.

A better merge function is also provided

https://gist.github.com/anonymous/8f468972d6286f403318c94bc6dbe382

from faceswap.

The face alignment script seems to rely on PyTorch, which is not fully supported on GPU for Windows users at this time. This could be blocking if we want it to be accessible and fast for non-tech users.

from faceswap.

Yes and it is also heavy as hell....

I'm still investigating because if it used just for faces landmarks, face_recognition does it pretty well also...

from faceswap.

Closing this issue because the referenced scripted has evolved and should not be integrated as is in the project.

I'll reopen an issue to handle the new script when I'll be more available on the project

from faceswap.

Hey guys @Clorr @Ganonmaster just accidentally found this :) I have recently addressed some accuracy issues and now the code should be on par (or very close) with the results reported in the paper.

Regarding support on windows, while official support will be added soon, there is already a conda package available that works fine on windows: https://anaconda.org/peterjc123/pytorch

With respect to the speed, the slowest bit is dlib face detector, however, one can replace it with almost any other face detection method.

from faceswap.

Hi @1adrianb , and welcome around!

Thanks for the info. Meanwhile we switched to face_detection lib since we are just using 2D landmarks and that face_detection was already used for face extraction.

If you think we can take advantage of any part of your work, feel free to suggest or even make a PR, we will take care of it with great pleasure!

from faceswap.