Teeny Tiny Pictures, II

[da]

This is a continuation of my previous post about photo mosaics, a project of Fish and I.

What do you get when you cross Flickr, GD Graphics Library, and a bunch of perl code?

In our case, something like this:



That is made from a corpus of 10,000 thumbnail photos from Flickr's "colorfields," "urbanabstracts," and "flowers" pools. Each tile is partitioned into a 5 by 5 grid and each of those is indexed by colours (RGB currently, HSV to come). The resulting data is stored as a 75-dimension vector.

We reduce the source image to a smaller size, such as 250x200, and each 5 by 5 grid is turned into another 75-dim. vector, which is compared with each of the tile vectors in the corpus. The closest vector is added to the image, subject to a few constraints such as not repeating nearby. A 5 by 5 grid works surprisingly well. Here is the top-left portion:



I love how the diagonal white line on the left travels across the tiles.

I assumed there would be mostly awful matches, cause y'know, what's the chance you've got an image with the proper colours in most of the parts of the tile? But the vector comparison choses the best fit overall, and 10,000 photos help. We're working on getting that to 30,000 or more.

This is the original photo, which I took in a huge indoor market in Budapest from melted_snowball's and my trip in 2003.



We're using Flickr's public API and only grabbing thumbnails, each only 1 or 2k, so it's not even a resource-hog. And the mosaic generation is quick- this image took me about 30 minutes to generate on an 800mhz computer.

We could've used ImageMagick as many of the other simple mosaic programs out there do, but we liked the interface of GD much better and it does everything we need for this. So- thanks boutell!!

Comments

[epi-lj.livejournal.com]

You know, if you can generate those in sufficiently high resolution, I have access to a very large colour plotter.

[chezmax]

I imagine you'll find the results will be even more remarkable in another colour space (I would probably suggest YUV or YCbCr over HSV./HSL)

Also, then, depending on your priorities, you can put more weight on the Luminance channel (which using, say, the JPEG equations for the colour transformation for YCbCr, would come automatically, as Y tends to be much larger than CbCr).

[chezmax]

Oh, in case this didn't come through: that's awesome. :)