Slashdot Mirror
The Algorithm That 'Sees' Beauty In Photographic Portraits
KentuckyFC (1144503) writes "Beauty is in the eye of the beholder. But what if the beholder is a machine? Scientists from Yahoo Labs in Barcelona have trained a machine learning algorithm to pick out beautiful photographic portraits from a collection of not-so-beautiful ones. They began with a set of 10,000 portraits that have been rated by humans and then allowed the algorithm to "learn" the difference by taking into account personal factors such as the age, sex and race of the subject as well as technical factors such as the sharpness of the image, the exposure and the contrast between the face and the background and so on. The trained algorithm was then able to reliably pick out the most beautiful portraits. Curiously, the algorithm does this by ignoring personal details such as age, sex, race, eye colour and so on and instead focuses only on technical details such as sharpness, exposure and contrast. The team say this suggests that any subject can be part of a stunning portrait regardless of their looks. It also suggests that "perfect portrait" algorithms could be built in to the next generation of cameras, rather like the smile-capturing algorithms of today."
Now all pictures will tend to be the same with the algorithm telling the amateur photographer how to frame the shot.
Never answer an anonymous letter. - Yogi Berra
The good news for many of us is that that the visual qualities of the person in the photographic have little impact on the beauty of the resulting portrait. “We find that race, gender, and age are largely uncorrelated with photographic beauty,” Instead, many of the factors that do correlate with beauty relate to the technical quality of the image. “Aesthetic score is related to sharpness of facial landmarks, image contrast, exposure, homogeneity, illumination pattern, uniqueness, and originality,”
So, no worries, here.
Slashdot, fix the reply notifications... You won't get away with it...
How would the algorithms deal with that?
https://www.google.com/search?...
We should be teaching machines to help us, not emulate our flaws and biases. beauty and ugliness, rich and poor, black and white, these are all concepts that in technology hold no relevance or meaning because theyre metrics by which we categorize and qualify people in often arbitrary and flawed ways.
Good people go to bed earlier.
i know some asian people and their idea of beauty is different than european. but for us europeans, this is old news. most models used on magazine covers and beauty queens have had similar facial shapes and features for decades. i remember seeing feminists on Maury and other talk shows in the 80's complaining how the media had a narrow view of beauty.
just as well as I do.
I've seen the classifications of other algorithms and they fall prey to the distortions of makeup. For example an average looking woman with makeup accentuating features, particularly dark eyebrows and eyelashes, will score higher than a beautiful woman without. The fact that the score can be affected by image quality suggests it's using the same sort of NN features. Better features for beauty are symmetry, facial width to height ratio, unblemished skin, bone structure, etc. ANNs tend not to understand morphology without deliberate preprocessing (e.g. by someone deciding it's a network feature worth calculation), so many of these things can be missed.
Proof that photos and real life can and do differ?
Hmmmm.... Imagine that. Congratulations on rediscovering what art discovered centuries ago.
This idea that the technical details of an image have a large affect in how we perceive the subject of the photo has been well understood for decades, even if it's not be well practiced by your every day "point and shoot" photographer. Just walk into even a low end portrait studio and think about how they all use the same kinds of lighting, flash, backgrounds, focal lengths and narrow depth of fields etc. There is a REASON they all look generally the same and it's not because they share the same studio designer, but that it's how you take good looking pictures. Now we have all this taking place digitally, with digital enhancement tools to take a so-so image and altering it to make it seem like a better one.
But even this is not new and predates the development of photography. The development of these ideas is evident if you follow the development of art though ages past. Go look at the more famous artworks and you will see similar characteristics develop in say painting, where we move from crude outlines on cave walls to near photo realistic efforts and beyond.
So us technical guys need to stop complaining about needing to take that "art history" course in college because there just *might* be something of value there, even to us... (smile)
"File to fit, pound to insert, paint to match" - Aircraft Maintenance 101
The trained algorithm was then able to reliably pick out the most beautiful portraits. Curiously, the algorithm does this by ignoring personal details such as age, sex, race, eye colour and so on and instead focuses only on technical details such as sharpness, exposure and contrast.
In other words, the computer was unable to learn to distinguish beautiful people from ugly people.
Don't waste your vote! Vote for whoever you want, unless you live in a swing state it won't matter anyways
From the FTA, three of the four most important features are both eyes and mouth sharpness. This should be no news to women, that have been using lipstick and eye-liners since forever.
Algorithms that result in a properly exposed and in focus picture? I think that's been done.
[...] then allowed the algorithm to "learn" the difference by taking into account personal factors such as the age, sex and race of the subject as well as technical factors such as the sharpness of the image, the exposure and the contrast between the face and the background and so on
Curiously, the algorithm does this by ignoring personal details such as age, sex, race, eye colour and so on and instead focuses only on technical details such as sharpness, exposure and contrast.
Something does not compute
Yahoo has labs???
Eventually it will be possible to prove so called modern 'art' ugly and rubbish. Great news.
The camera turns to reveal a large oil painting of a solid red rectangle.
It must have been something you assimilated. . . .
... is that by and large, there are some technical features that most humans - or at least most humans that the scientists from Yahoo Labs in Barcelona used to train this computer - agree create a beautiful image.
Knowledge is how to play a game, intelligence is how to win, wisdom is knowing what game to play.
Related: see also apps that detect whether a photo has been photoshopped. https://itunes.apple.com/us/ap...
-- I was raised on the command line, bitch
So photographers who aren't subscribers of the f/64 school of photography (to critically summarized: technical skill of taking and making the photo is what transforms a good photograph into a great photograph) are losers. I'll keep Nan Goldin, Cindy Sherman, Henri Cartier-Bresson, Garry Winogrand, Robert Frank, Man Ray, and thousands of other photographers who make excellent works of art but capturing the decisive moment (H.C.B. and Winogrand) or obtain a level of intimacy with and about their subjects (Goldin) or selves (Sherman) that any technical shortcomings are mere distractions to the gravity of their works.
Those photographers I've mentioned are quite "accessible"* to people outside of the fine art community, though nearly all of them do include nudes or nudity in some of their works, so are can be NSFW browsing, hence no links.
* By accessible I mean you do not need to have a comprehensive background in fine art or photographic criticism, history, or art theory to understand. They works are often considered to be appreciable by "outsiders," like myself.
This type of analysis would be better for taking pictures rather than analyzing existing ones, and even more so for outdoor photography (landscapes, natural events, etc.).
The best landscape/storm photos are about timing, you have to be there when the awesomeness happens and snap the shot.
Think the Grand Canyon full of clouds:
http://i2.cdn.turner.com/cnn/d...
Imagine a scenario where you setup a camera at a great view/vista (somehow securely) and leave it there for a while, letting a system decide when to take photos. Program it to capture sunset/sunrise (depending on the camera's orientation) every day, and stormy conditions (it sees lightning and then starts taking shots).
Let it take batches of photos that match some predefined conditions, and then go get the camera to look for the shot(s) that are awesome.
BlameBillCosby.com
To match human responses? Simple:
Table-ized A.I.
The thing is that people who take the most carefully lit, composed and focussed images (which is what the computer is using as it's metric) are professional photographers. They use models who are generally considered "beautiful". So unless you're very careful about your initial data set, you're going to come to some very bad conclusions...and that's what seems to have happened here.
So, this is bogus science...badly done.
-- Steve
www.sjbaker.org
As anyone who has experienced Dorothea Lange's "Migrant Mother" can attest, while the subject has never been described as classically "beautiful," the photograph is normally described as beautiful (and powerful) by almost everyone who has seen it.
My guess is what they've really determined is that:
1. Better photographers take better pictures, and also are more competent technically (ie. they take sharp, well-lit pictures)
2. People put more effort into getting technicals right when they're shooting something beautiful
Taking sharper photos of dull objects will only get you so far; the correlation is due to stuff that's deeper and harder to control: the subject and the photographer's skill/effort.
Let's not stir that bag of worms...
The best landscape/storm photos are about timing, you have to be there when the awesomeness happens and snap the shot.
Problem is, the photo you linked is from someone taking a snap.
Sure, if you're familiar with that vista when there aren't clouds then it conveys how unusual the scene is, but as a photograph it's nothing special at all.
A photograph like http://www.terragalleria.com/p... would be a better example. That doesn't happen because someone with a camera just went "oh, that's cool" *click*
A reworked version of last sentence:
"It's similar to how executives want the latest Powerpoint version in order to not look like all the other Powerpoint presentations floating around."
Table-ized A.I.
there's the other regular unreadable hipster website link spammer
I was thinking of ways to capture the rarer moments without having to be present. The Grand Canyon full of clouds is pretty rare (I've seen much better photos than the one I posted).
The system would also have to have a configurable way to manage exposure settings when taking a batch of photos. This is accomplished via bracketing features (which would also lend itself to HDR photos).
There's a particular view I have taken thousands of photos at over the years, it's that type of location I'm thinking about.
The Sunsets section here are some of the photos I have taken at this location, which came to mind in this case (Fort Kaskaskia Historic Site in southern Illinois):
http://fortkaskaskia.org/Photo...
BlameBillCosby.com
If (RND(1)>0.5),"beautiful", "not beautiful";