8 replies to this topic

[Randy]

This is a pretty cool one. And shows that Google is apparently getting a lot better at being able to tell what an image actually represents than I'd previously known.

The Labs similar image thingee is here.

There's a video announcement of it from a few days ago here.

The cool part is that it allows you to refine the search in interesting ways. As mentioned in the video you can get Jaguar's or other animals by what they're doing. So you can find similar face shots, action shots, etc. One I tested that they didn't mention, just to see how good it really was or if it was just looking at words on the page was a search for "st louis busch stadium", without the quotes. You see they built a new baseball stadium a couple of years ago and tore down the old one. But both stadiums are in St. Louis and both are named Busch Stadium.

The similar images thingee can evidently tell the two apart! Which leads me to believe they're actually analyzing the characteristics of the images themselves, as opposed to simply relying heavily on nearby text on the page.

For instance if I click on the similar images link for one of the brick facade pics, which is the new stadium, I get no images of the old stadium. Same if I click on the open outfield similars that is the new stadium (you can see the Arch in the background) I get only the new Busch Stadium. And if I click on the concrete circular building (middle of the 2nd row for me) that was the old Busch Stadium I get mainly pics of the old stadium that's now gone. There don't appear to be as many of those so there is some fill in.

Pretty impressive!

[Gerry White]

It is cool - but apparently our old deputy prime minister (a job he had the accolade of being so good at, that they scrapped the post!) apparently looks like Rene Zellweger ... http://news.bbc.co.u...ine/8014247.stm I am kinda glad I am not famous enough yet for there to be enough images of me on the internet to be a victim of this ...

I think that de-duping image searches will be interesting, or being able to say I want that image, but bigger ... it has also been interesting to find out where people we have worked with (not clients of course) have stolen images from (cough, Wikipedia usually ... )

Gerry

[rolf]

yeah, had a play with this and I think it's a really good idea.

Simple idea, straightforward to use and from a provisional play seems to work so I could definitely see it taking off.

[NASA]

looks like keyword relevant alt attributes are getting more valuable even when not an anchor.

[gsimerlink]

Would this have anything to do with that program Google had a while ago where two users would suggest tags for a photo they are shown simultaneously?

I can't recall what it was called, but it was a great idea. You would be matched up with another people and you would be shown a series of photos and have to tag them with description keywords within a 5 minute block of time. At the end all of your matches would be attributed to the photo and you would earn points for matches. It would be hard to game the system since your partner was selected as a blind random. It seemks like this could have been the start for the similar image search.

[Jill]

Would this have anything to do with that program Google had a while ago where two users would suggest tags for a photo they are shown simultaneously?

I was thinking same thing! I wouldn't be surprised if that's how they gathered much of the info that they are now using.

[Randy]

Dunno. That might have been a starting point. But Google being Google, my guess is that they've automated as much of it programically as possible. Meaning something on the back end is actually analyzing each image pixel by pixel. Then sees if that matches anything in their already established categories.

[Yoshimi]

I seem to remember reading something a couple of years ago in new scientist about an algorithm being developed (not by google, but I wouldn't be surprised to hear they had bought it) to allow computers to recognise shapes and colour in the same way we do. The reasoning was that if you point at a chair and tell a computer it's a chair, the computer will associate just that chair with the word, but if you can get the computer to look at individual characteristics within the image as a whole, you can get iit to recognise other chairs.

I remember thinking at the that it would seriously change image search, and it seems it has done just that.

[NASA]

well I decided to try my own and it failed miserably.

Type 'chelsea' without the quotes, all I got was images of Chelsea Football Club, not one Flower, let alone Miss Clinton or a Current Bun!

definately WIP me thinks!