backtop

Print E-mail del.icio.us 13 comment(s) - last by HighWing.. on Oct 22 at 1:36 PM

  (Source: University of California, San Diego)
Researchers at the UC San Diego have put Google's "Sets" widget to good use by giving computers the ability to identify objects by context

Google's free widgets have been used in many applications other than those Google initially anticipated. For example, some organizations use Google Maps to pinpoint locations of certain buildings relating to that group's business. Even Google has applied mapping applications like Picasa to identify the location of the setting of a user's digital photo. Then there are the more advanced applications such as giving a computer the ability to identify objects in a picture based on context.

Researchers at the University of California, San Diego and the University of California, Los Angeles announced a breakthrough of "common sense" by creating an algorithm that uses a not-so-used Google widget called Google Sets. For those unfamiliar with Google Sets, the web application accepts up to 5 words by default. These words can be as related or unrelated as the user would like them to be. When the user clicks either "Large Set" (which returns greater than 15 items) or "Small Set" (which returns 15 items or fewer) Google Sets will produce a page of results with words that can relate to a combination of the words submitted by the user.

By utilizing Google Sets, the UCSD and UCLA researchers have been able to take a digital photograph of a woman playing tennis and give a computer the ability to identify the specific objects in the photo (PDF) such as a tennis ball, tennis racket, tennis court, and the person playing tennis.
In the scene of a tennis match, four objects are detected and categorized: “Tennis court”, “Person”, “Tennis Racket”, and “Lemon”. Using a categorization system without a semantic context module, these labels would be final; however, in context, one of these labels is not satisfactory. Namely, the object labeled “Lemon”, with an appearance very similar to a “Tennis Ball” is probably mis-labeled, due to the ambiguity in visual appearance. By enforcing semantic contextual constraints, provided by an oracle, the label of the yellow blob changes to “Tennis Ball”, as this label better fits in context with other labels more precisely.
In this case, the oracle would be Google Sets which would help pinpoint the theme of the photograph and choose what it believes the round yellow shape is based on that theme.

Though this algorithm used in conjunction with Google Sets may identify objects in a simple scene with a clean background like the image to the right, more complex scenes with busier backgrounds may prove to be difficult for now. Still, the use of Google Sets and other algorithms which share the same idea seems to be the right path to creating highly advanced image identification algorithms.

Comments     Threshold


This article is over a month old, voting and posting comments is disabled
Uhhhh...
By Alexstarfire on 10/19/2007 12:32:58 PM , Rating: 2
So how does this work exactly? We feed it a picture and it, the program, decides what are objects and what the context is?




RE: Uhhhh...
By Schadenfroh on 10/19/2007 1:14:52 PM , Rating: 3
Application:

Spammers use it to bypass antibot measures in which you have to type in what number (or number of objects) that you see inside a picture.


RE: Uhhhh...
By TomCorelis (blog) on 10/19/2007 1:59:04 PM , Rating: 2
This has some incredibly scary potential for law enforcement, I think...


RE: Uhhhh...
By Dactyl on 10/21/2007 11:06:03 PM , Rating: 3
Why would you be afraid of this system? It's useless.

If you want to fool it, just bring a lemon to your tennis game.


RE: Uhhhh...
By HighWing on 10/22/2007 1:36:43 PM , Rating: 1
quote:
This has some incredibly scary potential for law enforcement, I think...


How much more scarier is it then the fact that Casino's have had face recognition software for years to track when known scammers walk in? And recently police have been putting this in use on city streets for known criminals?

All this really means is now the software could then check if a person was carrying a gun.

Personally in this way I would be more worried about false positives.


RE: Uhhhh...
By Scorpion on 10/19/2007 2:23:59 PM , Rating: 4
From my understanding, just skimming through the paper...

They take an image, and pass into the system a list of object labels that are likely to be in the image. They are using a sparse training set trained offline to learn object features used for categorization. Their goal appears to be to support this training data by using Google Sets based on these labels, the system is able to analyse image results returned from a sparse google search, rather than a strictly defined training set, and identify the objects in the input image associated with these labels by on-line learned features. I could be wrong, but this is my understanding after a quick glance at the paper.

They're using some very hot recent CV ideas in their approach: BoF(Bag of Features), SIFT (Scale-Invariant Transform Feature), and CRF (Conditional Random Fields).


break through
By nayy on 10/19/2007 2:31:21 PM , Rating: 2
I really think this is a break through, but what if that was actually a lemon in the middle of the court. This type contextual analysis will be really bad identifying odd placed object. In their defense the human mind has the same weakness,
check out page 18 of the link below. I have seen a video showing this effect, but i couldn't find it.

http://books.google.com/books?id=Yh-whOSPKSwC&pg=P...




RE: break through
By Alexstarfire on 10/19/2007 3:25:10 PM , Rating: 2
I hope you find that because it would be interesting to see if I would recognize the stuff that is off.


RE: break through
By Webreviews on 10/19/2007 9:02:22 PM , Rating: 2
Nice link. Thanks.

I had heard about a project where people around the world type in capchas, and that these capchas are actually words that OCRs find hard to read. Humans are good at reading the capchas and thus are training the OCR software. Is this what is happening with the image labels.

Is Google actually utilizing groups of individuals from around the world to label objects in pictures that appear on Google Images? Is that where the labels originate from?


Google's Contribution
By GaryJohnson on 10/19/2007 2:24:55 PM , Rating: 2
From reading the .pdf, it sounds like there are many far more impressive technologies used in this system and Google Sets just handled the object error checking.




omfg...
By lompocus on 10/20/2007 3:58:48 AM , Rating: 2
wow...amazing, BUT...I saw the lemon, and thought...

what if the program they use develops a perverted personality that takes over the world and makes many nasty things show up when we type 'banana'?




AMAZING!
By Inkjammer on 10/20/2007 6:45:44 AM , Rating: 2
With a tool like this I'll have no problem finally finding Waldo!




the world is ending
By inperfectdarkness on 10/22/2007 8:52:38 AM , Rating: 2
yes, folks, it's true. now when you image search on google, you can find theme-based pron even easier!

the search engine will now specifically look for "DD's" and all other sorts of debauchery.




"There's no chance that the iPhone is going to get any significant market share. No chance." -- Microsoft CEO Steve Ballmer

DailyTech Poll
Which web browser do you use on your primary personal machine? 






44 Comments







botimage
Copyright 2009 DailyTech LLC. - RSS Feed | Advertise | About Us | Ethics | FAQ | Terms, Conditions & Privacy Information | Kristopher Kubicki