Loading...
When it comes to search on the web, does size really matter? Cuil has made claims at a bigger image index than Google, but I haven't been able to do any successful searches on Cuil yet. While size matters, it seems clear that quality is a pretty important factor.
In fact, I think that the quality of the results is much more important than the mere size of the index. For example, let's compare Incogna's image index with that of our fellow image search engine: Tineye. Our current index is only about 10% the size of theirs, but let's compare the quality of identical image queries in the two systems:
versus 
versus 
versus 
Now you might argue that I've cherry-picked those results. But those results were easy to find, I urge you to try it for yourself.
One of our key assets is that Incogna's technology generalizes the shape of each object, which means that the Incogna engine can link two completely different pictures of the same object, building, scene, etc. In computer vision, that's pretty difficult. Oh, and we're also quite scalable too.
Another site that has a relatively "small" index (with 200,000 images) yet returns excellent results is Pixsta. As it turns out, maybe size isn't so important after all...
Posted by Kris Woodbeck on 2008-09-11
Here is a quick demo showing the Incogna search engine in action:
Stay tuned folks, we're on our way.
Posted by Kris Woodbeck on 2008-08-11
It seems like Graphics Processing Unit (GPU) supercomputers are starting to appear everywhere. It's a quiet revolution at the moment, but one that will certainly grow to consume supercomputing as a whole.
The Argonne National Lab recently announced a GPU Supercomputer (capable of 50-75 TFLOPS) and there is also one being built in France (capable of ~190 TFLOPS on the 48 GPU nodes alone).
If you want to be the fastest supercomputer in the world (i.e. beat IBM's Blue Gene which can do ~596 TFLOPS), all you need is about 600 hundred GPUs. You would be able to outperform their 200,000+ processor setup in less than 1% of the space, and for much cheaper. But don't worry, your regular processor isn't about to become obselete. We still need those to do more simplistic text based processing. GPUs are better suited for visual tasks like processing images, video, rendering and other highly parallel tasks.
For Incogna, this has given us quite the supercomputing bang for our buck. The average desktop processor can do maybe ~5 GFLOPS. Incogna's first hardware purchase over 1 year ago gave us several thousand times more processing power than that. Talk about being in the right space at the right time!
We can't wait for the launch of the NVIDIA Telsa rackmount supercomputer. Viva la revolucion!
Posted by Kris Woodbeck on 2008-07-28
A question I get asked a lot is: "is image search even popular?"
My typical answer is that it's the second most popular Google offering. But, at a recent Google presentation, R.J. Pittman pointed out that image search is in fact very popular. In 2006, 360 million image searches were being done per month; but today, that number is hundreds of millions of image searches per day. This is, quite simply put, an astounding level of growth; yet the major engines are all being fueled solely on text meta-data.
Last year, I won an elevator pitch contest where I claimed that image search is in the same state as text search was in the mid 90s: everyone are meta-crawlers. Google knows that this needs to change, and they're trying.
The second most common question I get is: "will visual ads make money?" I'll let a quote from the presentation answer this:
As the use cases for image search broaden, so too does the relevance of commerce related activities. This is exactly what we saw in web search several years ago. There's a direct connection between these uses and commerce. - R. J. Pittman, Google
With this drastic increase in image search, displaying relevant visual ads alongside search results is clearly an area for significant growth in online advertising. But, the visual processing required to make ads relevant is no simple task. Success in this area will clearly reap great rewards.
Posted by Kris Woodbeck on 2008-05-29
I've been testing a new image search engine from Idee Inc, called TinEye. They are (also) a Canadian company and have been doing copyright-based image processing/search for a few years now. TinEye is in beta and is being called the "Google of Image Search"
Basically, TinEye is heaven for copyright holders. It's great for finding photoshopped versions of a file; if you want to know who is using your images, TinEye will tell you.
But the real question is: how useful is this for everyday image searchers? The benefit of Google is that it helps me find what I don't already have, whereas TinEye only helps me to find (alterations of) what I already have. The problem is this: I've used TinEye, but I still need Google Image search to find what I want.
Here are some areas for improvement:
In any case, TinEye is a fantastic utility for copyright holders and it's great to see healthy image search competition out there!
Posted by Kris Woodbeck on 2008-05-13