Slashdot Mirror
Interwoven Patents Some Aspects Of Image Search
prostoalex writes "InterWoven patented locating and identifying image content via shapes, texture, color or resemblance to another image. No official word yet on whether the company thinks there are any infringers."
Google doesn't look at the image, just the filename, alt tag and surrounding context. Likewise with Ditto. I fail to see how that involves "shapes, texture, color or resemblance to another image". There are other companies out there that should be worried, but the ones you mention are about as far from that patent as you can get and still search on images.
These guys are a closer match, but since they are doing 3D CAD/CAM models, perhaps they are safe to.
On the other hand... these guys (eVe Image Search Toolkit) could be in trouble if they are not the patent holders themselves.
This patent seems more applicable to finding images that have similar color properties and gross image shape, which could be really useful when looking for images that go well together when compositing, not for finding pictures of a specific thing (unless you have an example that is very similar to the object you seek.)
So for the forseeable future, metadata will be far more successfull at finding images. Computer vision is still incredibly primitive: more so than computer speech recognition ten years ago.
Sig under construction since 1998.
I believe Google just uses surrounding textual metadata from the web page to identify a likely candidate image... combined with a false positive approach where an image named or labeled money.gif on a page with 50% content about puppies.. probably the image isn't a puppy, while all other large file sized images most likely are puppies...
A fool throws a stone into a well and a thousand sages can not remove it.
I'm not one to suggest violence but at the rate the patent office is going we Americans might have to walk up to the front door, kick it in, and beat some F%&*#&% sense into those people.
-=[ Who Is John Galt? ]=-
Honestly, if this company has an implementation of this system that works reasonably well, they deserve a patent on it. Extracting shape/texture information into (I'd assume) feature vectors effectively isn't the kind of trivial "development" that a system that performs different functions depending on how long a button is held down is.
However, if all they are patenting/developing is the searching, they're douchebags. I say this because after you have the feature vectors, the next step is a Nearest Neighbor Search, and there are already a number of algorithms for determining nearest neighbors. Unless their method somehow gets around the "curse of dimensionality", or provides other major improvements, I will be unimpressed.
I have discovered a truly remarkable sig which this margin is too small to contain.
The patent is more concerned with the method of converting the image into a manageable dataset that can be searched. So, it does not seem to rely on feature vectors at all. Moreover, this does look different than approaches I've seen to do the same thing. Fractal compression of images isn't terribly new, it's covered a bit in "Chaos and Fractals: New Frontiers in Science." However, I am not familiar with this method of classifying images based upon results from such a technique, this may very well be a novel development... dare I say, even worthy of a patent?
I have discovered a truly remarkable sig which this margin is too small to contain.
... but some of the more interesting stuff happened more recently (but still awhile ago). I have recently been looking at Fast Multiresolution Image Querying as a means to find similar images for f-spot. But it sounds like this patent is very broad and generalized, and systems like those described existed long before the patent. In fact, the paper linked above describes some of those systems.
GMail invites for completed freeipods.com of
Based on a short read of the article and the patent, I think that my former company (Audre, Inc., now eXtr@ct and Carnegie Mellon's Robotics Institute, among others, produced prior art as far back as the middle 1980's, back in the day before software and algorithms were patentable. I would go so far to say, at least at first glance, that some of the claims are superceded by convolution hardware dating to the 1960's at GE.
GE generated X:Y location & degree-of-match for an image regarding each of a set of simple image filters (rings of various radii and angular slits at 2 degree separation). This data was then cross correlated (some experiments used early neural nets, IIRC). They were successful at finding different types of features in aerial photography, such as farm, urban, water, grass, and forest.
The Audre Entity Recognition system used, among other things, input from a convolution/correlation system and a variety of other feature extraction methods, and used various means to build feature models from scanned engineering drawings, contour maps and other large format images. The system could produce a complete 3D terrain model from a simple contour map. The Visual Understanding Lab at CMU with which we worked also worked on using color features, more than Audre did. We even explored X-ray images, but scanning hardware of the time didn't have sufficient reliable gray scale capability.
A company in Denver or thereabouts was building systems using fractal decomposition of images as the fundamental data model for both display and recognition. They used a hexagonal cell model rather than the common rectangular one.
The patent is written in "patentese", so it'll take some study before I can be sure.
[Easter Egg: Check these movies (1, 2) and animated gif of ray tracing at 0.99c, by R. Thibadeau at CMU.]
It's easier to be a result of the past, but more fun to be a cause of the future! http://www.spacefinancegroup.com/