Probabilistically Placing Primitives

As a follow-up to the stippling project, I worked with my supervisor Wolfgang Heidrich and lab-mate Lisa Streit to place primitives probabilistically on a blank canvas. You can grab the resulting paper on my publications page.

Goals

The goals are to generate interesting drawings, reminiscent of what an artist might draw. If the artist was extrodinarily drunk or hit on the head several times with a large, blunt object. Many people are working on this problem, and we are looking at a specific set of restrictions:

Image-based input: still pictures, animations, etc. In particular, this means no 3D geometry for pasting things onto.
Minimal human input: we should be able to throw a picture at the algorithm and receive a drawing as output, tweak-free.
Real-time output: around 30 frames per second, because interactivity is good.
Flexible output: as many different styles as possible.
Attractive output: this seems to be an unquestioned goal in computer graphics. I'm not sure how often we achieve it, however.

Caveats

The caveats are many:

Any real artist would laugh at the overall quality of our output. Teaching a computer about shape or form, let alone mood or meaning, is extremely difficult. We cannot replace human artists, but we can steal tools from their toolbox.
Image-based methods will have frame-to-frame coherency problems in general. 3D geometry-based methods will be able to rely on the coherency of the graphics hardware. Our method, being image-based, definitely trips up on these coherency issues.
We are ignoring colour, since we don't have proper theory to cover it.

Examples

Here are some examples of the output we are generating:

Fonthill Abbey	Bee	Bust of Beethoven
Human foot bones		Fluffy Stanford bunny
Cezanne's "Still Life With Apples"	The ubiquitous Lena	Bust of Max Planck
Tiger	Van Gogh's "Self Portrait With Felt Hat, 1888"