Few questions about the paper

[Owen-Fish]

Thanks for your excellent work!

I recently saw this article on arxiv and would like to ask you for some details.

1. At the "Background Transfer" part. there has a sentence like:

We then aggregate each pair of corresponding features (fis, fti) by replacing the components of fti for the inner-face region with their counterparts in fis. I want to know how this region-wise replacement do?

2. Landmark encoder use the "Encoding in Style" methods, and your input to this network is landmak points (eg. 68*2 ) or landmark picture(eg. picture plot with landmark)? how two inputs can generate face shap lantent direction $\overrightarrow{n}$.

Looking for your reply, many thx!

[YueWangpl]

For 1., according to the Decoder code:

  def forward(self,source_fea,target_fea,mask):

        # for x in self.decoder:
        #     print (x)


        for i in range(self.log_size-2):
            mask = transforms.Resize(source_fea[i].size(2))(mask)
            blended = source_fea[i] * mask + target_fea[i] * (1 - mask)

            if i == 0:
                inputs = blended
            else:
                inputs = torch.cat([blended, outputs], dim=1)
            #print (i,inputs.shape,self.decoder[i])
            outputs = self.decoder[i](inputs)

        return outputs