lattas/AvatarMe

Question about the RealFaceDB

07hyx06 opened this issue · 2 comments

07hyx06 commented

Hi, thanks for sharing the dataset.

Is AvatarMe trained on the RealFaceDB dataset?
The paper says the original dataset contains 200 subjects, each with 7 expressions. Is RealFaceDB the anonymized version of the original dataset? Does it mean the 512x512 patches in RealFaceDB are sampled from the 200 subjects? What about the diversity of the RealFaceDB dataset, is it the same as the original Light Stage dataset?

lattas commented

Hi, yes, AvatarMe is trained on the same dataset. The patches are randomly sampled. They have also been cleaned from patches that show clearly identifiable features, as per the capturing license.

07hyx06 commented

Thanks for your reply! Can I ask some more questions?

  1. What is the UV topology of the RealFaceDB? I notice that AvatarMe and the ++ version are built upon GANFIT. I checked the GANFIT paper and found that it uses the Large Scale Facial Model (LSFM). Does the RealFaceDB use the UV topology of LSFM? Is the LSFM the same topology as the Universal_Head_3DMM as I find the FitMe project uses this universal head model?
  2. Will the code to balance the dataset's skin tone (mentioned in AvatarMe++ and FitMe) be released?

Thanks in advance.