Earlier this week, a journalist and photographer Esther Honig came up with an interesting project to see what beauty standards were like in different cultures. She took a selfie and asked photoshopers around the world to "Make Her Beautiful". The results are fascinating. Each photoshop touch-up exhibits a different kind of beauty standard.
Image credit: Esther Honig
However, with a closer look, I found that these variations might not only come from the ideals of beauty, but also come from different levels of photoshop skills. This inspiring project makes me wonder: is there a better way to transform Esther's selfie into beautiful women around the world?
My attempt is to use the readily available data on the web. We assume that the average faces of women in one country is the ideal face of beauty. To get the "typical" faces in different countries, we downloaded the average face of women around the world from the internet.
Image credit: Collin Spears
To better see their differences, we detect the facial landmarks and align these images according to their eye positions, as shown below.
Similarly, we detect facial landmarks from Esther's selfie.
Since we did not transfer the appearance (e.g., skin color) due to lack of image samples, it might a bit difficult to spot the differences. Here are a few samples to illustrate the smooth transitions.
If Esther were from ...
With the shapes from Esther's selfie and the average face of women, we can now morph Esther's selfie toward various styles. Here are the results of add 50% of the shape from the average face of woman. (Click on the photo to see the image at higher resolution.)
So what's the differences among the average faces of women around the world? We can study this problem using the shape and the appearance of these average faces.
By warping all the faces to their mean shape, we can get the "average face of the average faces". In some sense, this is the uniformly mixed beauty from 40 countries.
Next, we visualize what's the appearance variations among these faces. Here, we warp all faces to their mean shape so the only things that are different are the skin tones.
By combining the shape and the appearance features together (as did in the Active Appearance Models), we can compare the visualize the similarity among these faces. We use the Locally Linear Embedding (code available here) to embed these high-dimensional features to the two-dimensional embedding space.