Found Art: Model Predictions of Visible Distortion
This is one funny project if you are used to looking at low-rez digital images from an aesthetic point of view. These computational neuroscientists from NYU are working on a model that can predict whether or not an image will appear distorted to the human eye. They don't talk much here about why this kind of modelling is valuable, but I can imagine there are a quite a range of possible uses, not the least of which being a step in the ongoing project of technologically simulating the human brain. But the surface logic of the demo (to a lay observer like me) is so back-asswards that it's extremely cute.
There's a jpeg of Einstein, pretty good resolution. Then another that's somewhat compressed and a third that's super pixelated. These images are presented to the model, which simulates some of the neurological stages of vision but mashing different kinds of data sets. See a graph & explanation of the model.
The results are blurry jpegs! Yay! And these blurry jpegs tell us whether the original jpegs will appear blurry or not. SO cute. It's totally not an art project, but if it was it would have all the self-reflexive tautology of 70s high-conceptual art combined with the artist/activist DIYtechnology-for-DIYtechnology's-sake aesthetic of online art in the 90s. See the results. |
As someone who has spent a lot of time looking at compression artifacts (so as to better bring you shows like "Canada's Next Top Model",) I find this pretty fascinating, actually. To have a metric which gives a realistic prediction of image crappiness is pretty valuable stuff.
Hm. That is interesting. But if you already have the image, what good is the prediction?
I like to be told what I am looking at.
BTW Greetings from God's country. Literally. I am in Stockwell Day territory, but fortunately lots of vineyards. Am drunk.
Well, certainly you "already have the image" when you are testing it.
But let's say you wanted to run several video streams on the same pipe, and dynamically adjust the compression ratio for each stream, to optimize bandwidth? Compression artifacts are highly dependent on scene content and motion, so if you had something that could, say, watch the feed and allocate more bandwidth to the channels that started looking like crap, and have a reliable way to measure "looks like crap" you'd be pretty happy.
I see what you're saying now -- thanks for explaining, Rob. I should make it clear that I really don't mean to deride the computational model or the scientific goals. The stuff that's funny is only funny if you switch up contexts and pretend it's an art project - which it is not!!
|
This is one funny project if you are used to looking at low-rez digital images from an aesthetic point of view. These computational neuroscientists from NYU are working on a model that can predict whether or not an image will appear distorted to the human eye. They don't talk much here about why this kind of modelling is valuable, but I can imagine there are a quite a range of possible uses, not the least of which being a step in the ongoing project of technologically simulating the human brain. But the surface logic of the demo (to a lay observer like me) is so back-asswards that it's extremely cute.
There's a jpeg of Einstein, pretty good resolution. Then another that's somewhat compressed and a third that's super pixelated. These images are presented to the model, which simulates some of the neurological stages of vision but mashing different kinds of data sets. See a graph & explanation of the model.
The results are blurry jpegs! Yay! And these blurry jpegs tell us whether the original jpegs will appear blurry or not. SO cute. It's totally not an art project, but if it was it would have all the self-reflexive tautology of 70s high-conceptual art combined with the artist/activist DIYtechnology-for-DIYtechnology's-sake aesthetic of online art in the 90s. See the results.
- sally mckay 9-10-2010 3:03 pm
As someone who has spent a lot of time looking at compression artifacts (so as to better bring you shows like "Canada's Next Top Model",) I find this pretty fascinating, actually. To have a metric which gives a realistic prediction of image crappiness is pretty valuable stuff.
- rob (guest) 9-10-2010 6:23 pm
Hm. That is interesting. But if you already have the image, what good is the prediction?
- sally mckay 9-11-2010 12:53 am
I like to be told what I am looking at.
- L.M. (guest) 9-11-2010 6:50 am
BTW Greetings from God's country. Literally. I am in Stockwell Day territory, but fortunately lots of vineyards. Am drunk.
- L.M. (guest) 9-11-2010 6:53 am
- sally mckay 9-11-2010 4:00 pm
Well, certainly you "already have the image" when you are testing it.
But let's say you wanted to run several video streams on the same pipe, and dynamically adjust the compression ratio for each stream, to optimize bandwidth? Compression artifacts are highly dependent on scene content and motion, so if you had something that could, say, watch the feed and allocate more bandwidth to the channels that started looking like crap, and have a reliable way to measure "looks like crap" you'd be pretty happy.
- Rob (guest) 9-11-2010 4:55 pm
I see what you're saying now -- thanks for explaining, Rob. I should make it clear that I really don't mean to deride the computational model or the scientific goals. The stuff that's funny is only funny if you switch up contexts and pretend it's an art project - which it is not!!
- sally mckay 9-11-2010 5:22 pm