Researchers Develop 3-D Search Engine

An anonymous reader writes "Researchers have developed new search engines that can mine catalogs of three-dimensional objects, like airplane parts or architectural features. All the users have to do is sketch what they're thinking of, and the search engines can produce comparable objects."

Depends too much on

[Anonymous Coward]

the user's ability to sketch. I used to teach an amature art class, most people can't even draw a recognizable hand.

pr0n

[JohnGrahamCumming]

And they developed this for searching for "industrial objects". Riiighht.

Seriously though the pr0n industry is an extreme early adopter of most technologies, I'm sure that the researchers could fund research for the rest o their lives by creating an adult search engine.

John.

Re:pr0n

[theM_xl]

Are you capable of drawing any female pornstar so that she looks different enough from the others that the search engine can find her? More important, would you actually CARE if it came up wrong? :D

Any neural net people here?

[Thinkit4]

I've been curious if there has been a hyperelegant neural network solution that can do everything from recognize handwriting to faces. Completely different applications would only require retraining. So you wouldn't hear about specific applications like this, but one breakthrough that can simply, elegantly be applied anytime recognition must be performed. Are there any real (not patent or such) related reasons this doesn't exist?

Re:Any neural net people here?

[nacturation]

I've been curious if there has been a hyperelegant neural network solution that can do everything from recognize handwriting to faces. [...] Are there any real (not patent or such) related reasons this doesn't exist?

Aside from the fact that about a third of the brain (if I remember correctly) is dedicated to visual processing and that the brain has a hundred billion neurons, each neuron firing at up to 1KHz and with thousands of connections to other neurons? :)

It's an enourmously complex task to be able to do this reliably. Recognizing a straight-on photograph and matching it up with a corresponding mugshot is a whole lot different than seeing the side of a person's half-shadowed face from slightly behind them and recognizing that as the same Alice who held up the convenience store on Tuesday.

Handwriting recognition is orders of magnitude easier to do, since it's a lot easier to recognize similarities between a sample of writing and previously analyzed samples. In fact, there's a lot of research into this [ibm.com] already. Facial recognition is slowly getting there, but people are still stumbling over the same mistakes that were made decades ago -- attempting to formalize facial recognition by defining a set of rules and matching to those rules. Much like spam filtering, this works to a degree until the differences between the one you want and the one you don't no longer fits within the rules.

Neural net + genetic algorithms/programming to refine the net's connections and behavior is probably a good approach. Finding a means of populating, storing, and computing hundreds of billions of nodes in the network is the real challenge I think. Of course, take everything I've said with a huge grain of salt as this is a subject I'm fascinated by but have little practical experience with.