| Author | Thread |
|
|
05/15/2009 05:03:04 PM · #51 |
Bah... Random Nbr... both of these images got an identical 26.4
 
Dunno about you but I know which one I like :-)..... |
|
|
|
05/15/2009 05:03:16 PM · #52 |
I checked my last 3 Free study entries
 -84.6
 -94.5
 -80.6
And my Best of 2008 winner
 -62.8 |
|
|
|
05/15/2009 05:31:40 PM · #53 |
92.4
 |
|
|
|
05/15/2009 05:36:43 PM · #54 |
I've been working in computer vision for 17 years and I've never before come across any researcher with the audacity to claim their software could predict aesthetic quality of photos. There is plenty of solid work in this field on predicting specific responses to specific types of photos (like predicting which faces we find most beautiful), but predicting the overall "aesthetic quality" of a photo is just plain nuts. I skimmed through one of their papers listed on the "About" page on their site, and I have to admit that their overall approach seems sound. But I still don't think I would have approved the paper if I'd been on that review committee, they just make way too many major simplifying assumptions.
They extract 56 different types of features from each photo and run them through a classifier which has been trained on thousands of rated images from photo.net. The biggest problem is that designing many of those features is an extremely difficult problem, and they seem to make huge simplifying assumptions. Like their Rule-of-Thirds feature, which they claim measures whether "the main element" in the photo is near any of the four thirds-line intersections. But as far as I can tell, they don't even try to determine what's "the main element", instead they just compare hue, saturation, and brightness around all four intersections to hue, saturation, and brightness everywhere else. There's no way this tells you how well an image fits the rule of thirds! I could go on, but you get the idea...
Edit: Reading the other two papers they list on their About page, they didn't just use photo.net photos, they also used 16,509 DPC photos! No wonder their system doesn't work ;)
Message edited by author 2009-05-15 17:46:57. |
|
|
|
05/15/2009 06:19:34 PM · #55 |
Originally posted by magnus:
Edit: Reading the other two papers they list on their About page, they didn't just use photo.net photos, they also used 16,509 DPC photos! No wonder their system doesn't work ;) |
I don't recall them asking permission to do that ... :-( |
|
|
|
05/15/2009 07:08:59 PM · #56 |
Originally posted by magnus:
they also used 16,509 DPC photos! No wonder their system doesn't work ;) |
Makes sense. A large base of easily accessible photos that have already been rated according to aesthetic appeal. Made their research easy. Now I don't feel bad about loading down their servers with our own experiments today. |
|
|
|
05/15/2009 08:55:24 PM · #57 |
| Hmm.. we've been guinea pigs. That might explain my recent cravings for sunflower seeds and raw veggies. |
|
|
|
05/17/2009 10:24:56 AM · #58 |
//acquine.alipr.com/acquine.php?mine=1&start=0&like=vote
So far my average is 35 for 143 image and four images are over 90%
Message edited by author 2009-05-17 11:07:22. |
|
|
|
05/17/2009 10:50:31 AM · #59 |
It's much, much worse than a voter whose taste differs from mine. At least people tend to have some consistency in what they like.
This thing might as well be a random number generator.
 : 1.7
 : 14.7
 : 37.2
 : 84.7
Message edited by author 2009-05-17 10:54:27. |
|
|
|
05/17/2009 11:08:56 AM · #60 |
Originally posted by magnus:
I've been working in computer vision for 17 years and I've never before come across any researcher with the audacity to claim their software could predict aesthetic quality of photos. There is plenty of solid work in this field on predicting specific responses to specific types of photos (like predicting which faces we find most beautiful), but predicting the overall "aesthetic quality" of a photo is just plain nuts. I skimmed through one of their papers listed on the "About" page on their site, and I have to admit that their overall approach seems sound. But I still don't think I would have approved the paper if I'd been on that review committee, they just make way too many major simplifying assumptions.
They extract 56 different types of features from each photo and run them through a classifier which has been trained on thousands of rated images from photo.net. The biggest problem is that designing many of those features is an extremely difficult problem, and they seem to make huge simplifying assumptions. Like their Rule-of-Thirds feature, which they claim measures whether "the main element" in the photo is near any of the four thirds-line intersections. But as far as I can tell, they don't even try to determine what's "the main element", instead they just compare hue, saturation, and brightness around all four intersections to hue, saturation, and brightness everywhere else. There's no way this tells you how well an image fits the rule of thirds! I could go on, but you get the idea...
Edit: Reading the other two papers they list on their About page, they didn't just use photo.net photos, they also used 16,509 DPC photos! No wonder their system doesn't work ;) |
Interesting and you share pretty much the same opinion I have. I spent the last couple of days experimenting with this - one image, but I altered it in Photoshop and got anywhere from a 10 to a 97.4 just messing with saturation and hue. Brightness didn't seem to be much of a factor at all. The system seems to be partial to reds and brighter shades of green. You want to try something really interesting, use blank colors in a frame coverted out to .jpg and submit those. :>)
There's also another interesting effect. I can vary the "asthetic" quality of an image by simply compressing the image. I started by submitting a 50K image, then jacked it up to just under 2 megs (which is their limit on submissions). The "asthetic" value changes accordingly. Try that with the blank color and you'll see that it's total bullfeathers as my Maternal Grandfather said in polite company.
Later,
Tom |
|
|
|
05/17/2009 11:09:24 AM · #61 |
Originally posted by GeneralE:
Originally posted by magnus:
Edit: Reading the other two papers they list on their About page, they didn't just use photo.net photos, they also used 16,509 DPC photos! No wonder their system doesn't work ;) |
I don't recall them asking permission to do that ... :-( |
Seems a bit cheeky doesn't it.
Later,
Tom |
|
|
|
05/17/2009 11:23:30 AM · #62 |
Originally posted by Yo_Spiff:
Originally posted by magnus:
they also used 16,509 DPC photos! No wonder their system doesn't work ;) |
Makes sense. A large base of easily accessible photos that have already been rated according to aesthetic appeal. Made their research easy. Now I don't feel bad about loading down their servers with our own experiments today. |
True, but the problem is (1) permission to do so and (2) the variable and extremely subjective nature of what is valued from an "asthetic" standpoint.
Later,
Tom |
|
|
|
05/17/2009 03:59:22 PM · #63 |
 =97.5 :)
I think their system needs some modifications LOL |
|
|
|
05/17/2009 04:07:02 PM · #64 |
My second-highest rated image  (and a Staff Pick over at Jones Soda -- go vote for it!) got a 4.2 -- just about my usual score here too ... :-( |
|
|
|
05/22/2009 08:00:16 PM · #65 |
Crazy! 6.4632 here. 96.9 / 100 on their system. |
|
|
|
05/22/2009 08:17:22 PM · #66 |
Just voted. Gave it a 4.2.
j/k
ETA: Looks like it's 8+ over there.
Message edited by author 2009-05-22 20:19:08. |
|
Home -
Challenges -
Community -
League -
Photos -
Cameras -
Lenses -
Learn -
Help -
Terms of Use -
Privacy -
Top ^
DPChallenge, and website content and design, Copyright © 2001-2025 Challenging Technologies, LLC.
All digital photo copyrights belong to the photographers and may not be used without permission.
Current Server Time: 08/02/2025 04:26:47 PM EDT.
