Hey all, a new paper just came out which claims to be the new “state-of-the-art” in “Face Beautification” - I think this paper is extremely problematic since it seems like it is trained on an extremely biased dataset and mostly it just performs “Face Whitefication” (see Figure 5 especially).
I was wondering what do other people think about it, and most importantly how do we make it clear as a community that we outright reject such unapologetically racist and biased results?
@rachel - sorry for tagging you directly but thought you might be interested in this.
I work in computer vision. I see more and more of these papers for automatic Beautification. I stopped thinking they were racist. Instead I think they are understandably myopic.
For prototypes, engineers work with the data of the community they are in. For production, if they want their work to have worldwide appeal, they need to diversify their data to suit the various communities. For instance, this paper appears to work with a company Oben, Inc that is targeted towards Chinese consumers (I believe). It makes sense to assume this is how Beauty is viewed among the community of interest. For worldwide appeal, they would need to change their dataset and how they define beauty.