Wow. With PhotoSketch, you just draw a sketch, label each item, like so:
and then the system goes out, finds photos that match the sketched items and their labels, and automatically pastes it all together into one composite image:
The site is down right now but the paper is available for download and this video gives you a taste of how it works:
Again, wow. (via migurski)
Update: I’ve seen many references to Photosketch saying that it has to be fake (here’s a sampling). But it’s pretty obviously real. For one thing, here’s the source code; try it out (Windows only). It was presented at SIGGRAPH Asia 2009; here’s the listing of papers presented. The authors all have web pages on university sites and have published work using similar techniques and technology (Ping Tan and Ariel Shamir for example). And is what it does really that unbelievable? At the most basic level Photosketch is just find me a man that’s sorta shaped like this, a dog that looks like this, and paste them together with a background that looks like this. That the results are so impressive (especially for a demo) is a testament to the team’s execution and attention to the small details. Even if it turns out to be an elaborate hoax, I have no doubt that someone could actually build a working version of Photosketch…I mean, look at TinEye and Photosynth.