Friday, April 08, 2005

Video Blogger

There's been a lot of speculation around what Larry Page, a big guy at Google whose title I don't remember, said about Google accepting user video submissions later in the week. It's ranged from the cool (video blogging?) to the yawn-worthy (provide the content for Al Gore TV?) to the intriguing (test content for Google to use to try to make searching video as easy as searching text?).

While it sounds like it is just the AlGore_ thing, it would have been awesome if it were video blogging. The options for showing Rachel's antics would have been great, and it would have mattered less if the lighting stinks when shooting in our house. Although you know there would have been a million posts the first day that started out "Stardate 20050408" and then went on to be painfully awful to watch.

Someone did uncover a project at Oxford code-named "Video Google." In the simplest form, you take a frame of video, draw a box around an object in that video, and then the computer will find other examples of what you drew. For instance, they had it find other instances of Bill Murray's tie in Groundhog Day, or a poster in Run, Lola, Run. (Best stuff is on the "How it works" page.) They even had a demo you could play with. It's pretty neat, but not entirely there yet. They have one old back and white movie you can play with, so I tried some guy's nose. It found a number of noses, but also fingers and a few other things. I also tried a guy's hat and I think those were even less successful. But the frames were in black and white and rather small, so sometimes it was hard to see what it had drawn boxes around on each of the screens it returned in the results. However, as fledgling technology, it shows promise and looks kind cool.

I know that Australia has had technology for years that allows a computer to read license plates from video. They use it to time how long it takes vehicles to get from point A to point B. They know how long it takes when you're doing the speed limit and if a semi passes point A and point B in less time than that, instant ticket.

But that's just looking for rectangles and then doing character recognition on it to get the licence plate number. A lot more "known variables."
Post a Comment