timofter

Please check our research papers linked on the webpage.

Our data consists only from normal (or natural looking) face images in the wild (from IMDB, Wiki, and/or BLINQ user profiles). On such data we get very good apparent age prediction (better than the human reference) and also very good gender and attractiveness prediction.

The attractiveness is highly subjective and its perception varies from one culture/region to another. We used data from Switzerland.

Our solutions are far from being perfect and the guessed results should not be taken too seriously.

We consider to update our models to explicitly deal with distortions and non-human face contents.

The ratings are "like" (positive) and "dislike" (negative). If men have ratings close to uniform distribution (44.81% are "like" ratings, 55.19% are "dislike"), women with only 8.58% "like" ratings are clearly unbalanced, or if you want the women are "very picky". We call this a strong bias in the ratings of women while for men there is a small(er) bias and this only to point out the difference from the uniform distribution (50%-50%). We can not tell and we are not the ones to judge if the men are voting randomly or are less picky or if the women are right or more picky, since the attractiveness is subjective and we do not have an ultimate ground truth to compare with. We can not say if the attractiveness should follow a particular distribution, we work empirically. In our study, the ratings of men tend to agree more and correlate more with our visual representation based on the face image (looks). We removed from our data the extreme cases such as users with less than 10 ratings or with too many ratings.

Hi sethbannon,

I am Radu, one of the authors.

We thought at similar experiments, however psychology is not our expertise. If you check our paper on hotness/attractiveness (http://www.vision.ee.ethz.ch/~timofter/publications/Rothe-ar...), I am sure that you'll find some interesting results on how different age-grouped people rate, a paradox, and more. And yes, there are many interesting experiments to do and questions to answer.

Hi tsumnia,

I am Radu, one of the authors.

After 2 years we face almost the same issues, but probably we cope differently with them. Note that our solutions are fully described in the two papers mentioned on our faces.ethz.ch page. For attractiveness we used data from Blinq.

Our apparent age estimation solution is the winner of the latest LAP challenge, ICCV 2015.

The first processing step consist from (human) face detection. We use the standard OpenCV for our faces.ethz.ch demonstrator. A failure of this step is likely to propagate in the unreliable/wrong attractiveness prediction. For attractiveness, age, and gender prediction we start from a cropped image assumed to contain a (roughly aligned) face as found by the detector.

I hope that this helps to understand the aforementioned result.

Yes, it is a foreseen application.

Hi fatjokes,

I am Radu, one of the authors.

As mentioned in our article (http://arxiv.org/abs/1510.07867v1), the training dataset for attractiveness is from Blinq. The underage and above 37 years old face images were discarded. All people are heterosexual in our training dataset and mainly from Zurich area, Switzerland. Therefore, our model of attractiveness fits the cultural bias of Zurich, Switzerland.

We consider faces.ethz.ch a little fun tool. I hope that the fans of Yoona will understand :) With more training data from abroad Switzerland our algorithms will fit their expectations.

For age and gender we used much more and diverse data, therefore are more reliable for the majority of ethnic backgrounds.

Yes, this was the data we worked with for attractiveness modeling. The algorithm learns from millions of ratings from males to females and from females to males, all heterosexuals. No underage or above 37 years old, and mostly people from Zurich area, Switzerland. For age and gender prediction we use more and diverse data.

I am Radu Timofte, one of the authors.