Wednesday, May 27, 2009

AQUINE = photo.net?


(William Eggleston 4.7)


I came across a flurry of posts recently about the AQUINE system - a not terribly good acronym for Aesthetic Quality Inference Engine - which is supposed to give "intelligent, unbiased and instant assessment of photos".

It is described as a machine-learning based online system for computer-based prediction of aesthetic quality for colour natural(??) photographs and is also supposed to shows that computers can learn about and exhibit "emotional responses" to visual stimuli like humans do.

Unfortunately, however, it seems right now that the folks at Penn State seem to have based it on the "photo.net algorithm". If you compare its top rated photographs with the top rated on photo.net, it's pretty hard to tell them apart. Among other things, lots of overdone HDR will get your picture a good rating...



(Atget 5.0)


Bearing in mind that Ctein found at least one pretty substantial flaw in it - that if his linked photos had a frame they scored dramatically higher than if they didn't - I decided to through it few spin balls to see how it rated some of my favourite photographers. Which means for one thing I was throwing it a good few B&W images rather than colour.




(Walker Evans 42.7)


As I had guessed, most didn't do too well - poor old Atget on got about a 5 for one 12 for another, and Egglston's tricycle got about the lowest at 4.7 (that should please a good few of the folks on APUG). Most were somewhere in the 30's or 40's - Walker Evans, Struth, Friedlander, Sugimoto. After that (yes, I know they aren't photographs exactly - well, photographs of paintings), I tried Picasso, Van Gogh - again, the poor things only scored around 20 or 30. Although Turner's 'The Fighting Temeraire' - voted the most popular painting in Britain got a 70.0.



(Van Gogh 22.7)


The only three I did get with high scores were Sudek at 91.8 (not surprising when at his most romantic, plus it has a nice black frame) and, a little more unexpected, Andreas Gursky who got 85.8. Lynne Cohen also got 87.0 - but that one also had "nice" colours in it.




(Andreas Gursky 85.8)


Now I wonder, as it is supposed to learn (and I have almost no understand of the computing aspects of this kind of artificial intelligence), that if a concerted effort was made to flood it with Eggleston, Struth, Parr, Graham etc etc photos, would it start to learn and become biased towards a sort of late 20th century New Color aesthetic instead?

But for now, if you want to work out where your work stands on a sort of 1980's Photo Club aesthetic scale, I think this is the place to go.



(Sudek 91.9)

5 comments:

Luis said...

A special note of thanks to Tim for running "Ringers" through Aquine.


Even if this thing worked to perfection, it is as meaningless as a Photo.net "Originality: 5.6, Aesthetics: 4.4."

How is that useful to anybody?


This is a pathetic addition to the photographic landscape's dystopiary.

How long before a dumbed-down version of this finds its way into DSLRs to tell you how to frame and when to shoot something that will please the masses? And if someone doesn't get it, you can always show them the score.


Imagine the poet saying to her beloved: How do I love thee? Let me count...ah... a 6.8.


--- Luis
_______________________________


Elizabeth Barrett Browning (1806-1861)

How do I love thee? Let me count the ways.

I love thee to the depth and breadth and height

My soul can reach, when feeling out of sight

For the ends of Being and ideal Grace.

I love thee to the level of everyday's

Most quiet need, by sun and candle-light.

I love thee freely, as men strive for Right;

I love thee purely, as they turn from Praise.

I love thee with a passion put to use

In my old griefs, and with my childhood's faith.

I love thee with a love I seemed to lose

With my lost saints, --- I love thee with the breath,

Smiles, tears, of all my life! --- and, if God choose,

I shall but love thee better after death.

Lee said...

My god sometimes i wonder why i get out of bed

qtluong said...

AQUINE learns to rate by associating features with a set of already rated images. The training set *was* photo.net. The features used were "camera-club level", like rule of third, saturation, texture. There is no way it would account for higher-level considerations which would apply in the images you tested.

Chris Rust said...

Hi, thanks for this. Although ACQUINE is an interesting project I share your doubts and I've posted a critique of some of the scientific aspects at http://chrisrust.wordpress.com/2009/05/16/acquine-is-this-how-to-engineer-affect/
This feels like a group of engineers who are trying to do something that might be worthwhile in the very long run, but their experiments are flawed by a complete failure to confront the question of what aesthetics actually means.

hellophotokitty said...

I'm with Luis...

DSLR's are already turing new photographers into zombies, and seasoned ones into lazy blind visual commentators. Sites like this make me think twice about the craft, which is quickly becoming, as noted, an pretty algorithm.