Slashdot Mirror
What Makes a Photograph Memorable?
Hugh Pickens writes "Researchers have developed a computer algorithm that can rank images based on memorability. They found that in general, images with people in them are the most memorable, followed by images of human-scale space — such as the produce aisle of a grocery store — and close-ups of objects. Least memorable are natural landscapes. Researchers built a collection of about 10,000 images of all kinds for the study — interior-design photos, nature scenes, streetscapes and others, and human subjects who participated through Amazon's Mechanical Turk program were told to indicate, by pressing a key on their keyboard, when an image appeared that they had already seen. The researchers then used machine-learning techniques to create a computational model that analyzed the images and their memorability as rated by humans by analyzing various statistics — such as color, or the distribution of edges — and correlated them with the image's memorability. 'There has been a lot of work in trying to understand what makes an image interesting, or appealing, or what makes people like a particular image,' says Alexei Efros at Carnegie Mellon University. 'What [the MIT researchers] did was basically approach the problem from a very scientific point of view and say that one thing we can measure is memorability.' Researchers believe the algorithm may be useful (PDF) to graphic designers, photo editors, or anyone trying to decide which of their vacation photos to post on Facebook."
tits
Boobs.
Exposed breasts
Appearance of Forrest Gump is undoubtely a plus.
Achille Talon
Hop!
An unfeasibly large anus.
Confucius say, "Find worm in apple - bad. Find half a worm - worse."
they fscked that study up real good.
personally I remember a really good sunset, or an awesome view of nature far more than I remember a friggin grocery store. heck there are two that I go into regularly and one I rarely go into and the one I rarely visit I can't find anything in it.
human sized spaces aren't more memorable. what it contains and that means to the viewer is what is rememberable.
i thought once I was found, but it was only a dream.
Boobs. Lots of boobs.
"Service unavailable" is burned into my head forever.
Cute, pink-haired cat-girl.
Please don't tell me that entities selected through Amazon's Mechanical Turk pass for subjects in MIT-level scientific research. Should I start taking MIT less seriously?
This study done by 80 year old grandmas that drag out their fucking cameras every time family gets together for any damn reason yet if they take a trip around the world it's just head shots of who they traveled with in front of nondescript backgrounds.
What a strange study. From the summary, the researchers sampled a group of individuals by presenting them with random photographs and rating not their _memorability_ but rather recall in asking them to press a key "when an image appeared that they had already seen". This is much different than what I believe makes a photograph memorable--which typically involves some sort of an emotional response to the subject in photograph. For instance, nature pictures taken on a journey to me personally would be very memorable--even though the study suggests otherwise.
If you're in marketing and want people to "recall" your product, yeah sure, this study is relevant. But, it's sort of misleading labeling it memorable as it suggests an emotional response and this study does not address that.
By the way, the definition of "memorable" is the quality of being worth remembering--very different from recall.
Of course things with people and animals (or representations thereof) are more memorable than landscapes. Our minds have evolved to put greater emphasis on things that are a threat or opportunity. Besides, landscapes are generally classified as that only because they're outdoors and don't have any other distinguishing characteristics that would put it into another group.
Knowledge Brings Fear
getting out of a pool, ...
"Researchers believe the algorithm may be useful to (...) anyone trying to decide which of their vacation photos to post on Facebook"
WTF? Do people really need an algorithm to decide that? Post what's interesting or what you WANT people to see... duh!
-Boll
If that photo wasnt isolated, but associated with external ( sound, light, smell, etc) or internal (i.e. feelings, idea associations, complex toughts, etc) things, getting back those things will make easier to get back those photos.
They should consult with more photographers. One thing is obvious: the most-memorable pictures have a central point of focus...something to grab your interest. The least memorable images in the TFA have nothing to grab your attention. That applies to a mixture of subject matter as well as a single subject, such as landscapes.
The TFA gave short shrift to aesthetics, too--where in the photo the central point of focus most favorably may be placed, such as the Rule of Thirds and Golden Sections. These go back to Da Vinci...not new ideas.
An algorithm is of mild interest, yet Van Gogh's Starry Night, a landscape, is far more interesting than his Starry Night Over the Rhone, a painting that includes a man and a woman walking arm-in-arm by the river.
Said that one thing that makes a photo interesting is a pic of something common in one location, that is shown in another location where that thing is not common. There is no way an algorithm could describe that.
I am Slashdot. Are you Slashdot as well?
Other criticism of the study aside, a group of people who might be interested in how well pictures are remembered after short glances are advertisers and marketers.
Easy. Scarcity. Pictures used to mean more back when we didn't have hundreds of them. That one lone photo of a vacation somewhere represented the entire trip. You would look at it and it would dredge up all the memories. Now you have pictures, postcards, even video of a trip. There is just so much of it that you never even look at most of it again. Is it really any wonder it isn't as memorable?
I address this question in a (to be published) book on the psychology of entertainment where I explore the concept of novelty. Although mere newness is not enough to make something memorable, if something combines a strong design structure or a vivid one, and is both personally and culturally novel, its memorability is greatly increased. When we are young (immature experientially), almost everything is novel and gets consideration as we take in perceptions. Repeated patterns in the environment are assimilated into recognizers, so that we can detect what is unusual and possibly a threat. (Ie. that which is out of place invokes attention, leading to better chance of survival from potential threats.) I believe that the same mechanisms, with varied parameters, then serve multiple purposes including artistic perception. mechanics of reading, and so on. I am engaged in an ongoing effort to embed this principle in hybrid symbolic and neural recognizer systems, as part of a larger effort. Anyway, I leave the take-away point that memorability is a function of both perceptual system operation and interpretive deep systems drawing on culturebases, hence novelty and memorability is dependent on individual (per person) frameworks.
Any of them ironically labled..
"Its not my weiner."
http://www.jsonline.com/blogs/news/123109063.html
-Hack
Got Geometrodynamics? Awe, too hard to figure out? Too bad.
This seems very applicable to marketing and commercials. Now they can engineer a way for their product to stay stuck in my head.
Boobies! And no, not the jiggly kind. I mean the bird, you perverts.
Enlightenment is a pipe dream. So where's the pipe?
This is not the first its kind out there. In Siggraph 2008, a paper claiming to measure how beautiful a face is. In Eurographics 2011, a paper claim to measure how good a facial makeup is.
The goal of such research is always to quantify something subjective (such as beauty, memorability, etc.). The general pattern of such kind of research is to first gather data through some kind of user study, then extract some potential features that could be used to explain the data, and lastly use machine learning to determine which feature or features best explain the data, then publish! The method looks scientific, and it seems our day-to-day intuitions can be explained based on scientific models. It is also very hard to argue they are wrong. There is no way to evaluate their result unless you do a separate user study, which most of us lack the time and money to do. Since the presented results match our intuition and make sense, these research are often given good reviews.
However, I want to raise the question: What is the point of such kind research!? This research does not give any new information except make us aware of our intuitions. Does it confirm our intuition is correct? No. Does it explain why our intuition is the way it is? No. Can the scientific model used in the general case? In this case, can an advertiser really use this algorithm to pick an ad that is more memorable? I doubt it. My reasons are the following
1. The data are noisy.
2. The study simply reveals a correlation, not causation.
3. The model is quite limited: many important things are not taken into account such as image composition, familiarity of the scene, etc.
4. The features that could explain the data are hand-picked by research, which means it is perfectly possible for a different set of features to explain the data just as well or even better, but those features are not tested in this research.
5. Machine learning typically only yield an over-fitted solution due to repeat usage of the testing set.
It might be a bit harsh to say such research is a waste of time and money. At least for me, it seems like just a different kind pseudoscience.
Satisfies none of the listed criteria, does it?
Oddly, these Internet memes don't seem that bad, with the mellowing of time. Here is what I think of them now...
tubgirl: blow up dolls are funny
lemonparty: reminds me of Barney Frank
goatse: what is the medical term for that?
Eh, YMMV.
This issue is a bit more complicated than you think.
caused a buffer overflow
Table-ized A.I.
You have never seen one of these?
http://www.anseladams.com/category_s/4.htm
Error 503 Service Unavailable
Service Unavailable
Guru Meditation:
XID: 258827006
Varnish cache server
Looking at the images, there's definitely familiar objects helping recognizability, but I saw a much more computationally-easy pattern. Look at the edge detection results of the images. Now count the internally terminated edges (of some length at-least 5px on the thumbs) vs the "lines" that extend to the end of the image.
Memorability = (internally-terminated edges) / (edges touching border)
At a glance, this appears much more likely. Why? I think people put more importance on fully-framed "objects" as life demands a high ability to detect objects and we do so with corners. The more unterminated edges, the less we can trust that we see "objects" rather than just data, so it's thrown-out as noise.
The formula should be easy to test:
1. Edge detect
2. "Walk" edges: remember ones that don't split, end, or considerably change direction (circles ok) in 5px (appropriately scaled)
3. Of all the #2 found edges, mark those that touch edges
4. Memorability = [ #2 (all) - #3 ] / #3
Let me know how it goes...
memorability Looking at the images, there's definitely familiar objects helping recognizability, but I saw a much more computationally-easy pattern. Look at the edge detection results of the images. Now count the internally terminated edges (of some length at-least 5px on the thumbs) vs the "lines" that extend to the end of the image.
Memorability = (internally-terminated edges) / (edges touching border)
At a glance, this appears much more likely. Why? I think people put more importance on fully-framed "objects" as life demands a high ability to detect objects and we do so with corners. The more unterminated edges, the less we can trust that we see "objects" rather than just data, so it's thrown-out as noise.
The formula should be easy to test: 1. Edge detect 2. "Walk" edges: remember ones that don't split, end, or considerably change direction (circles ok) in 5px (appropriately scaled) 3. Of all the #2 found edges, mark those that touch edges 4. Memorability = [ #2 (all) - #3 ] / #3
Let me know how it goes...
Science & open-source build trust from peer review. Learn systems you can trust.