I created an interactive, online version of the Felzenszwalb & Huttenlocher segmentation algorithm.

Here are the details of the engineering behind it:

Efficient Graph-Based Image Segmentation
Pedro F. Felzenszwalb and Daniel P. Huttenlocher
International Journal of Computer Vision, 59(2) September 2004.

The program takes a color image (PPM format) and produces a segmentation with a random color assigned to each region.

The parameters are: (see the paper for details)

sigma: Used to smooth the input image before segmenting it.
k: Value for the threshold function.
min: Minimum component size enforced by post-processing.

Typical parameters are sigma = 0.5, k = 500, min = 20.
Larger values for k result in larger components in the result.

Also, I reimplemented the average color algorithm

For the pseudo-physics engine, I was looking for a way to eval() expressions. This isn’t built into ActionScript 3.0, but there is an EcmaScript 4 parser/interpreter in the Tamarin project. Pretty cool. Now I have to figure out how to feed the results of eval back into the swf.

I chose to mess around with the pseudo-physics package this evening. I probably should have done work, or slept. Probably slept.

Anyway, I cleaned up the code…er, all the physics stuff is in one place. More importantly, there are controls for the equations.

The green box in the top right sends parameter info to trace. No use to you.

The dark blue circle is friction. Friction greater than 1 means things accelerate on their own accord.

The translucent purple dot is the black-hole, center of the universe.

You’ll figure the rest out.

[Pseudo Physics Simulator]

Ron Maimon and I worked on this flash animation program the other night at Stella’s. I really like it and want to do more. Half the fun was changing the pseudo-physics equation.

[Click image to view animation]

The charlesdietrich.com website has gotten a new look. I expect to asymptotically approach the ideal template, so there are future changes in store. However, I think this is a big improvement. Though not as big as an improvement as the previous change, from the WordPress default theme to the Connections theme…ah, asymptotes!

The current theme is equiX, with a few customizations. There’s a blog that I visit sometimes that uses this theme, so I hope I don’t look too similar.

The image is a current screenshot from my research project…I’ll probably use a photograph later, like I did for Julie (I designed the website, but the excellent photo is courtesy of Luke Walker).

It actually does something! Well, right now it displays a graph of the music and you can navigate around in the music. What was fun about this was I used a Model-View-Controller (MVC) architecture, which worked quite well. So well, that I transitioned the music animation program to this system.

I am really happy about the blogging thing. Anyhow somebody named Kokila Dhamodharan from India got in touch with me last month about hierarchical agglomerative clustering of images. It happens I know a thing or two about that. Basically you have an n-dimensional, metric space where the dimensions are properties of the image that you are interested in (color, variation, etc). When I did this I was comparing the partitions of a segmented image so I had color, size, shape, major axis etc.

I realize that the asymptotic complexity analysis of my paper was a bit slapdash. But who cares when you have run times of a second. I know that everything is less than O(n^3) anyway.

So here’s the naive approach: You calculate pairwise distances O(n^2). You have n^2 edges, and run an MST algorithm, which takes, with m=n^2, O(m log m) = O(n^2 log n).

Of course a really asymptotically efficient MST algorithm is essentially O(n). And you can approximate the set of pairwise distances — somehow or other…

Anyway, my code relied heavily on Pedro F. Felzenszwalb and Daniel P. Huttenlocher’s code, available here. I even reused their MST implementation. My project also used some code for an image format called ppm.

I sent my code over to the guy in India, but I am too lazy to really figure out what the license is, so I’m not posting it. It looks like it is MIT license compatible. Pedro’s new repackaging is GPL.

It would be fun to put either Pedro’s app and/or my app online.

Here are the results from my code:

MusicXML is a XML language for music scores. I have been playing around with it and composition software for several months. I have been using some composition software (Finale) that can generate an MP3 from a score. I have a few scores of classical and modern pieces. Perhaps I don’t have the instruments right, or something else, but the result is full of cracks and often sounds like dire organ music — it’s a lot of fun to listen to. I have a recording of ostensibly the same piece of music, and the two sound nothing alike.

MusicRain is a cool online score viewer / score player written in flash.

I needed a mp3 cue point creation widget written as a standalone ActionScript 3.0 project, so I started working on one. It occurred to me that once I get the sound graph built, and the sound navigation built, somebody might want to build in some editing functionality. So I put the project up on Google Code as as3soundeditorlib.You can see the demo here.

Here’s the screenshot:

Welcome to the newest page on the web (at this instant). I am working with ActionScript 3.0 on a few projects, some TOP SECRET. But never fear, I have an open-source project to create a standalone sound editor and cue point generator in ActionScript 3.0. I’ll share my techniques and code, and point you to good ActionScript 3.0 resources on the web.

Charles a.k.a “Toby”