PhotoSketch - makes photographs out of rough sketches

[KingDavid8]

According to authors, their software can take any rough sketch, with the shape of each element labeled with its name, find images corresponding to each drawn element, judge which are a better match to the shapes, and then seamlessly merge it all into one single image.

PhotoSketch's blending algorithm analyzes each of these images, compares them with each other, and decides which are better for the blending process. It automatically traces and places them into a single photograph, matching the scene, and adding shadows.

PhotoSketch

David

[jw]

Whilst I'm sure this will be improved on, I don't think photoshop editors will be too worried just yet, as the latter images didn't look up to much to me. The depth of field, orientation and size matching will be very hard to get right. I can see the user having to search a lot for the right blend of images.

This could end up looking like a free babel fish translation, which turns coherent sentences into random words.

Also, I wonder how they'll deal with any issues of copyright? Probably not.

ETA: I suppose Jenn is going to be a lot busier!

[Syllavus]

I'm dubious that this software is real. I know that imaging software can be capable of some amazing things, but as someone who uses Photoshop on a daily basis, I'm skeptical that this program is anything more than fiction.

[SpaceRobot5]

I'm skeptical of this program as well. I'll wait until they can provide an actual demonstration of the program before I actually believe this works.

They claim to be pulling these images off the internet. They could use a system like Flickr's which would have all the images tagged. The part I'm skeptical about is the matching of the sketches to pieces of the images. In the video the show the program finding and cutting out the frisbee out of a picture of a dog and a frisbee. It somehow matches that piece of the image to a badly drawn semi-oval line. That is some crazy computer-vision AI right there.

Though it may be simple for a human, that is an incredibly complex problem for a computer. Being able to cut a human shape out of a background and match it to a crude-stick figure drawing is pretty amazing. So amazing in fact that I'm disinclined to believe their claims until they produce some proof.

[Silas Sparkhammer]

Originally Posted by Syllavus

I'm dubious that this software is real. I know that imaging software can be capable of some amazing things, but as someone who uses Photoshop on a daily basis, I'm skeptical that this program is anything more than fiction.

I feel the same way, especially after seeing the example. That's definitely a case of "too good to be true."

Silas

[kjbrasda]

Not to mention the ethics of pulling images off the internet apparently without permission to use them.

[jw]

The Telegraph has this about it.

Looking up Crunchgear, who are usually reliable, I find this hyperbole statement,

"It’s early, experimental, and questionably useful, but it’s just too cool," writes Devin Coldewey on CrunchGear. "I don’t want to be premature here, but I’d say tentatively that this does appear to be the greatest thing of all time."

But he does go on to temper it somewhat with

They showed it at SIGGRAPH. If this works the way they say it does (and I doubt it does at this point, really), it’d be a party game all to itself. And can you imagine playing with this thing on a tablet? God damn! The future, people!

So there you have it. It is being worked on, by Chinese students, Although, I couldn't find any mention of it over at The National Basic Research Program of China, which is supposedly funding it. Also, this organisation seems to support an altogether much weighter list of projects.

However, my original comment is still valid. The images don't look anywhere near real enough to me. Imagine cutting out a bunch of magazine pictures and pasting them on another magazine picture, all kinds of stuff like shadows etc., would be missing. Unless they really find a way of altering the pasted images an awful lot of manual editing would have to be done to make these images look realistic, therefore defeating the purpose.

[SchmooPie]

It was presented at Adobe's Max Conference as an experiment based on theoretical algorithms. Nowhere near anything actually functional.
Here's a link to a Photoshop product manager's blog about it. The linked page also has a link to the research paper.

ETA- The video is a demonstration of how it might work. It doesn't purport to show working software.

[Insensible Crier]

Let's see some photographs generated for xkcd characters.