In this post, I will list down some tips for selecting anonymuos faces for anonymization.

Where to get anonymous faces (target faces)

The website https://generated.photos/faces provides a variety of artificially generated faces. You can signup for the website and there are really a lot of faces.

What (target) faces are good for FSGAN?

  • The faces provided by generated.photos are properly aligned and with plain background, which are ideal for FSGAN.

  • The face should have a slight smile. A no-smile face or big-smile may degrade the quality of output.

  • The skin-color of the target face is used. Hence, our face-bank should have wide ranges of skin-colors. Same for lip-colors, eyebrows and eye-color.

  • However, the face hair (beard, moustache or on-head) of target face is not used. Hence, it is fruitless to choose faces with variety of hair-looks.

  • Anything away from the target face or in target’s background won’t affect at all. Hence, the tag “Image by generated.photos” is not a issue for us.

  • You can also have aged people and kids.

  • The old people in generated.photos are not very old. Hence a near-old face is suggested for old face.

Special Cases

Special cases arise when there are several artifacts appearing on the face. The way to handle them is to have same artifacts on target face too.

  • We observed that it is more suitable to anonymize a face wearing glasses with target face wearing glasses. Similar for glasses with reflection.

Our face-bank

Link: https://drive.google.com/drive/folders/1EGiVI3fMLwNiYG-Es-Sy5qqie9Co2eZI?usp=sharing

We have the following categories of faces:

  1. black_female
  2. black_male
  3. white_male
  4. white_female
  5. old
  6. kids
  7. glasses

There are alteast 5 faces in each category.