Google Blogoscoped


PhotoSketch software turns line drawings into photos

Roger Browne [PersonRank 10]

Friday, October 9, 2009
14 years ago2,998 views

I think Philipp might find this software interesting.

It takes a sketch and some image search terms, and assembles a photo-like picture to match the sketch:

A video is here:

James Xuan [PersonRank 10]

14 years ago #

There's no demo, so I don't believe it for now.

Roger Browne [PersonRank 10]

14 years ago #

You can download and run the software. I won't link to it here, because it's in binary form only and from China so I'm not gonna run it, but you can find it by Googling.

Those who have run the software say two things. First, that it's "hacked together" and clearly a concept demo rather than an end-user tool. That's not surprising, because it's not touted as anything other than an academic research project.

Second point is that it doesn't grab its images from the internet at large, but from an image database which you need to set up. And it's not clear whether the researchers "tailored" their image database to get better results.

Nevertheless this software advances the "state of the art" by choosing images which match not only the search term but also the approximate sketched shape. From those, it chooses images that have a background texture similar to the texture of the image chosen for the overall background. Not only does this make it easier to merge the images seamlessly, but it gives better context (e.g. a dog playing on sand will look better against a sandy scene than against a grassy scene).

In the sample results, you can see some anomalies. The most common one I noticed is that the direction of the shadows in the foreground and background is often inconsistent.

Luke [PersonRank 0]

14 years ago #

there a demo video here:

Philipp Lenssen [PersonRank 10]

14 years ago #

Yes indeed I find this fascinating and had downloaded it some days ago. (I once pondered if it was possible to create an art collage via Google Images by simply labeling e.g. parts of a room to create a room still, or e.g. to create a face by googling eye, nose, mouth etc., but IIRC preliminary tests didn't turn up anything feasible... for instance, try googling for [right eye] returns too much variance to simply throw this together into a collage without further refinement.) It wouldn't run so I contacted one of the developers listed but they told me to search for a DLL or download some other graphics library, which I did but it still didn't work, so I will wait until this is available in some commercial release for now (someone might want to try and build an end user app for this idea, it's almost too good for that to not happen! Could they run this over CC-licensed Flickr images?).

Forum home


Blog  |  Forum     more >> Archive | Feed | Google's blogs | About


This site unofficially covers Google™ and more with some rights reserved. Join our forum!