Final Project: Recursive Mosaics

• Project Proposal

  This project will consist of two main parts. The first will be to implement automatic mosaic building. Some examples are included here and here. The basic idea is to build a low resolution representation of an input image using a library of existing images. This has been addressed in several papers (noted below). Using the Flickr library would be excellent for this purpose.

  The next part is where it gets fun! A suggested project was to create an artificial "infinite zoom". With the mosaic implementation from above, we can then "zoom" into part of an image. As the image enlarges we focus on one particular subtile, which in turn is built of it's own mosaic tiles. When stitched together into a movie clip, this should be quite interesting!

• Papers on automatic mosaic creation

  • http://portal.acm.org/citation.cfm?id=566633
  • http://www.cs.princeton.edu/gfx/proj/mosaic/princeton-cs-tr-574-98.pdf
  • http://charity.csfac.uwlax.edu/StudentResearch/Documents/ImageMosaics.pdf
  • http://ieeexplore.ieee.org/iel5/10196/32539/01521085.pdf?arnumber=1521085

• Preliminary Results

  • 
    6500 source images in 64x64 pixel tiles
  • 
    6500 source images in 32x32 pixel tiles
  • 
    6500 source images in 16x16 pixel tiles
  • 
    6500 source images in 8x8 pixel tiles

• Super Extra Pretty!

  Per jmccann's suggestion, I took it upon myself to create an accelerated flythrough implementation. Using some of the starter code in the Neon Helium tutorials I was able to get a simple OpenGL project off the ground. The following is a simple demo of this. The video doesn't quite do it justice since the transitions are nice and smooth while the images remain large and of high quality in my application.

• Enhancements

  • Multithreading. In my 2-day hacked implementation of this, I didn't take care to multithread anything. As is, once a new slides is shown, ImageMagick is sluggishly trying to resize the source image in the background. This can lead to quite a long wait time before the mosaic appears. This should all happen in parallel.
  • Alternate zoom method. The current implementation scales each image linearly with time to create the zoom effect. This has the result of making it appear to start fast and end slow. I like it, but maybe a polynomial scaling would be more appropriate.
  • Tile/Target blending. In my original mosaic code, I do not blend the mosaic with any part of the target. While this is fine for images that are blocky and easy to identify, this doesn't look so good on most photographs with a lot of high frequency components.
  • Randomize tiles. It looks pretty odd when there are a lot of tiles with the same image. There should at least be some sort of variation. This can be solved by imposing some rules on tile reuse.