The PDF branches off of Active Shape Modeling (by Tim Cootes, https://secure.wikimedia.org/wikipedia/en/wiki/Active_shape_...) with a style I'm not entirely familiar with, however basing off my background I think I can expain a little of what is going one.
Active Appearance Models are a extension of Active Shape Modeling, effectively taking a set of landmark points on a shape (in this case a face), and averaging them out to the 'average shape' of the face. AAMs take it to the next level by taking this average face and warping the original landmarked face to the average shape. From here it takes the average of the textures of the face, producing eigenvectors. These vectors would be used as a unique identifier if the program was running some form of face recognition.
In lieu of face recognition something you can use the averaged AAM face for it recreating a face BASED off the average (say for this paper's example: beauty). By generating an averaged AAM model off only those subjects that scored an 8 (out of 10) or above, you create the 'average attractive face'.
Now, if you take John Doe's face with a score 5, and generate the eigenvectors that would represent his face based on the 'Good-Looking People Only Scale', it would create a better looking version of him.
Another application of this technique is digital face aging (example here: http://www.intechopen.com/source/pdfs/14645/InTech-Implicati...) in which instead of 'morphing' the subject's face to look more attractive, you morph the subject's to look 'older' based on statistical averages based on age range, gender, and ethnicity.
There was a reconstructive surgeon that did a lot research into beauty, because he had to remake faces given partial information. He found golden ratios for all the major features of the face. He constructed a transparent mask with guidelines and did wonders for his patients.
This was the best application I have heard for similar research. Not so much for the money aspect but for the good it did.
The issue with this comes into play of manipulating their face from different angles, which is still an issue in the biometric world. OKCupid's posted statistics on the 'more attractive' profiles, and not all the profiles use the stare directly into the camera shot.
What do you guys think about a smartphone app that could be panned around to capture a 3D image then used the golden mask to calculate an objective attractiveness score?
Would be some fun engineering and probably pretty popular.
then used the golden mask to calculate an
objective attractiveness score
Except that you can't calculate an objective attractiveness score, even if there is statistical consensus among raters. Might as well just throw some random numbers on the screen, it wouldn't matter much anyway since your app will be less effective and gather less interest than HotOrNot.com did 11 years ago.
A much better project would be a piece of software that automatically enhances the aesthetic appeal of people in photos. Imagine looking great in any profile picture you upload on Facebook (or whatever), without you or your friends noticing the change (blaming it on distortions from image resizing).
It would be a cool app, and allot of new techniques may be invented or discovered while developing it. But one should ask what are the attractiveness criteria? what do you mean by "Data-driven". The data comes from other users? I mean aesthetics norms comes from society or the user, or both?! Very subjective! if we build tools to enforce aesthetics norms on users from the "common sense" - (things that never made sense too me), we are basically killing all the peculiar beauties in the world, say, a cute pimple next to the lip.
Idea was that you could get an objective "hotness" or attractiveness rating based on how well your features fit the golden ratio / golden mask. It would also give users preferences when they first use the app (eg against 10-20 faces, which you can have precalculated all the points accurately), you can map each face against their preferred faces, giving the personalised rating system.
Use case at a pub/bar quick scan of the iphone would give a objective rating and your rating.
V2. could add location and start building real time tracking of where the hot women / men or those matching closer to your rating were going tonight :)
Generally, the code is very hacky, and (for example) the user may need to do a lot of steps to run it, many counter intuitive.
In the Mythical Man-Month, Fred Brooks says that to go from a prototype that does everything you want, to one with nice packaging, takes 3x the effort. I'm not sure about the 3x number, but it does take a lot of effort.
Most might not have the licensing issues. I'm not sure on the copyrights for things like Viola-Jones or Active Shape Modeling. I know Cootes has refused to release the source for his modeling software in case their financial incentive.
The software is out there, I'll link some if you want. What you are really looking for is the training data used to create the faces though. That is where the money is.
I have a feeling that's very common. I recall a big to-do about Cortexica and their brain-inspired computer vision algorithm. They built it into an iPhone wine label recognition app, which currently has two ratings; both one-star.
Close, but no cigar. According to OkCupid[1]―and data mined from its users―the most "attractive"―in this case, measured by messages from previously unknown users received―pose for men is not smiling and not looking at the camera, and for women is "flirting" and looking at the camera (although in both cases smiling gives "above average" attractivity).
I think that it's more important to look friendly, or approachable, to be attractive to someone. In my experience I've found that a smile is better than a scowl, and definitely better than a neutral expression.
True beauty comes from inside. My grandparents weren't terribly attractive... but I'll never forget my grandfather's smile. When he smiled it lit his face in such a powerful way I can see him even now (he's been dead almost 9 years!).
And I don't think that women looking "flirty" is attractive. Looks more predatory to me.
I don't think that women looking "flirty" is attractive.
But they are more attractive - for the same reason that women wearing short skirts are more attractive and it may not be for the same reason you think it is.
The reason has to do with accessibility: women, unlike men, are very picky when it comes to dating. We (men) know this and when a woman doesn't look like she's looking for men, we often choke or don't bother pursuing a date. A flirty attitude is the strongest indication that a woman is interested.
Looks more predatory to me
Some people, men or women alike, often end up with dumb looks in photos :: that's what happens when wanting to simulate sensuality when you're not in the mood, not to mention some people are just not photogenic and really need the help of an expert for such photos. I am talking about genuine sexy looks above, not "let me eat your brain" looks :)
Man, I'd hate to be one of the models for one of these pictures. "Ohhhh yeah," say thousands of internet users. "That computer-generated version really is a lot better-looking than the real one due to those horrible flaws on that person's face". No thanks, I'd rather not see the difference between my face and a better-looking version of my face, it'd just make me feel bad.
On a broader point though, the human taste for "average-looking" faces is an interesting one. Aside from anything else it's an interesting Nash equilibrium -- I want to find a mate with an average-looking face so that we can produce children with average-looking faces who will be the most desirable breeding partners in the next generation.
What I really wonder is whether "average-looking" is hard-wired in, or whether we're programmed to spend our childhood scanning all the faces around us and mentally averaging them out to determine what a human face should look like. I suspect the latter -- it's a much more stable strategy over evolutionary timescales, and also explains things like why people often find people of their own race more attractive, and why mixed-race people are often unusually attractive.
I'm too lazy to look up the ref, but someone told me that the preference for average faces was actually an artifact of the averaging process. It turns out that averaging images of faces also makes the skin look a lot younger and smoother, and therefore more attractive.
One way to test it would be to see if the median face, by some choice of metric, was also more attractive. The problem with other averaging methods, like simple pixel averaging, is that you can produce a result that is not actually particularly similar to any face anyone has. For example, if you had a population where everyone's face was highly asymmetric, but 50% in one direction and 50% in the other direction, the average would be a symmetric face that is completely atypical for the population.
How does the second hypothesis explains both of those tendencies at the same time? It would rather seem to predict that people of mixed race would be found less attractive then those of the same race as the subject and more attractive then those of the other race. I.e. an averaging out of attractiveness rather then an increase of it.
Oh, it depends where you're brought up. What you really find attractive is a mix of all the races in the community where you were brought up. If everyone you meet in childhood is of the same race then you'll probably find that race attractive, but if you (say) spend half your time hanging out with whites and half your time hanging out with Asians...
>Aside from anything else it's an interesting Nash equilibrium -- I want to find a mate with an average-looking face so that we can produce children with average-looking faces who will be the most desirable breeding partners in the next generation.
Isn't it a wrong Nash equilibrium? You should choose a mate with someone on the other side of "average", so your children will have average faces. An average face and your (more likely than not) non-average face do not make average faces.
This "other side of average" effect is to some extent seen with the major histocompatibility complex.
I don't think we have a preference for average looking faces.
We do have a strong preference for symmetrical faces, maybe because that indicates you've been very healthy for many years.
And averaging out a bunch of non-perfectly symmetrical faces, results in a very symmetrical face.
But that's very different from average. We strongly prefer decidedly rare features showing off high levels of sex hormones. An average face would have average sized eyes for example, but we prefer bigger eyes which indicate youth.
But I 100% agree with you on having a picture of your better looking twin, no thanks.
I would think the opposite, as the 'familiar face' idea would lead to more inbreeding in groups, thus less healthy offspring. Plus, one should keep in mind that for millions of years, beauty was not as important a constraint as proximity and availability.
Active Appearance Models are a extension of Active Shape Modeling, effectively taking a set of landmark points on a shape (in this case a face), and averaging them out to the 'average shape' of the face. AAMs take it to the next level by taking this average face and warping the original landmarked face to the average shape. From here it takes the average of the textures of the face, producing eigenvectors. These vectors would be used as a unique identifier if the program was running some form of face recognition.
In lieu of face recognition something you can use the averaged AAM face for it recreating a face BASED off the average (say for this paper's example: beauty). By generating an averaged AAM model off only those subjects that scored an 8 (out of 10) or above, you create the 'average attractive face'.
Now, if you take John Doe's face with a score 5, and generate the eigenvectors that would represent his face based on the 'Good-Looking People Only Scale', it would create a better looking version of him.
Another application of this technique is digital face aging (example here: http://www.intechopen.com/source/pdfs/14645/InTech-Implicati...) in which instead of 'morphing' the subject's face to look more attractive, you morph the subject's to look 'older' based on statistical averages based on age range, gender, and ethnicity.