Hello all,
these days I was reading some scientific articles dealing with photographic image aesthetics.
The idea behind is to develop software which will automatically divide images into two groups - good and bad.
All researchers are aware that this is extremely hard task, unsolvable at the moment.
Their test and train database usually consist of 10% of best and 10% of worst images downloaded from some
photo comunity site (for example 60000 images in total, for research they take 6000 best and 6000 worst
photos).
If they tend to make usable software, usable for an average user,what kind of database they should develop?
Maybe regarding some features this 6000 best images may be representative (color, contrast, sharpnes...), but
regarding some other features may not. For example - composition. An average user will follow (more or
less) rule of thirds, golden rule etc. and will rarely make good photos breaking these rules. Pro user will break
it more often making extraordinary photos.
My question IS what do you think about the before mentioned selection of images for the database (10%
of best and 10% of worst images)?! Is this selection representative or not?
It is clear so far that at the moment computer vision can not accomplish such tasks.
Regards, Matija. |
