Computers With Opinions On Visual Aesthetics 125
photoenthusiast writes "Penn State researchers launched a new online photo-rating system, code named Acquine (Aesthetic Quality Inference Engine), for automatically determining the aesthetic value of a photo. Users can upload their own photographs for an instant Acquine rating, a score from zero to 100. The system learns to associate extracted visual characteristics with the way humans rate photos based on a lot of previously-rated photographs. It is designed for color natural photographic pictures. Technical publications reveal how Acquine works."
Re:Rubbish. (Score:4, Interesting)
It rated one of the best pictures I've ever taken as 13. Then it rated a fairly generic cityscape at around 60.
I think it has some learning to do.
Re:Acquine may assign funny scores... (Score:4, Interesting)
I noticed this with a picture I took in France that everybody praises. It got 6.9. I did the smallest possible change in color, darkening it imperceptibly. The new version got 35.7. Doing a selective gaussian blur also tends to raise the result a lot.
My rating of their algorithm is 0.01 star, which can be summarized as "it sucks".
Interesting (Score:3, Interesting)
One thing first. There *are* certain esthetic and technical rules / guidelines which are what we could call "objective" in the sense that they are very general. For example, a photograph usually looks better if the composition is balanced, if the 2/3rd (or golden mean) rule is used, the lines in the picture are coherent and lead the eye in the right direction (e.g. towards the subject), if the photograph is correctly exposed, colors matched etc. Of course, some of the greatest photographs break those rules; however, like in many things, you succeed in breaking the rules if you know what you are doing, and you cannot do it very often.
I can imagine that you can come up with an engine that is able to detect how "rule conformant" a given picture is.
However, pure formal esthetic judgement is what we rarely mean when talking about a "good photograph".
There is one main issue that will make it very hard to match our "overall" esthetic sense. Firstly, we are unable to detach the image contents from the "pure form". That means, if we see a worried women holding a child, we cannot just look at that as a composition. Also, we are always considering what we know about the subject. E.g. if we have a photograph of a man standing in water, if the photograph ends just below the place that his legs go into the water, we will have the impression that his legs are cut off, and that there is something wrong about the photograph. Finally, facial expression is immensely important for the perceived esthetics of a photograph.
I did some experimenting -- some of the truly great photographs of our times got rather lousy scores (e.g. Dorothea Lange's famous photograph, but also some color photographs as well), while at the same time rather random shots I did of my sons got even five out of five stars. Well. Maybe it will still be useful to someone to filter out the worse photographs.
j.
Re:Pulitzer versus Goatse.... (Score:2, Interesting)
Well that's exactly the point, isn't it? We're a long way from AI being able to judge artistic merits, because doing so often isn't just a property of the image alone, but is related to what the things in the image represent.
Now having said that, it might be interesting to have some judgement that is not biased by cultural perceptions - one that treats Pulitzer and Goatse on an equal playing field, judging what they look like rather than what they represent. But it's unclear what meaning such an algorithm has - it's not comparable to what humans would think, nor is it meaningful to say it's the computer's own opinion (I mean, anyone could write an algorithm that assigns some arbitrary value based on the image data, but what use or meaning that has is another matter.)