Questions about the results in supplementary material about homointerpgan HOT 5 CLOSED

Thanks for being interested in this work.

1. Yes, the RaFD datasets are well aligned to the CelebA. The alignment tool will crop the head based on the head location of the reference image.

2. Only frontal faces are used.

from homointerpgan.

@yingcong Thanks for your quick reply!

I try the provided facealign/align_all.py, which it aligns the face to a near-frontal position and good synthesized results could be obtained from that.
But I am still wondering how to put the synthesized face back to the original image, as we have to deal with the scaling/ translation/ rotation problems? I am stuck there.
Could you please share the tool with us? Then we could test the method in a better way. Or could you give us some suggestions on addressing the problem?

Looking forward to your reply. Thanks.

from homointerpgan.

We would like to generate similar results as Fig.1 - 3 shown in supplementary material

from homointerpgan.

Hi @RegisWu, For blending the faces back, we basically leverage the "seamlessClone“ in opencv. You can refer to https://www.learnopencv.com/seamless-cloning-using-opencv-python-cpp for the usage.

from homointerpgan.

Note with thanks!

from homointerpgan.