An update

[gfish]

What am I up to, you may be asking¹? Well, classes are almost over. I have 2 more assignments to do, one huge and one pretty small. But mostly, now that I have a supervisor, I'm focusing on the 3 class projects due at the end of April. No finals this term, which is nice, but that means projects instead.

Image Understanding II: I'm implementing this paper, adding image features from several others, in order to create a generalized image quality assessment for the SRVC project. Basically, we want to screen out cartoons, paintings, illustrations and renderings, while at the same time ranking the remaining images by how good they are for training a classifier.
Status: Pretty good. I have the paper implemented, and I'm starting to add the new features. I have a decent training set downloaded, but it could use some more work.

Sensorimotor Computation: I'm implementing some visual tracking/servoing on the eye simulation hardware. Not entirely sure what, yet. Basically GYRE in one dimension, with a motor instead of compressed air thrusters. And no freefall.
Status: The person who really knew how to run the system left in December, so it took me about 3 weeks to even get it running again. But this weekend I started working on the new code, which so far hasn't been that hard. I need to implement some kind of PID control loop now.

Machine Learning: For reasons that are too arrogant to be written down, I want to find a way to match the states in two or more Markov processes as equivalent, given just the initial probabilities and the transition table. As far as I and the professor know, no one has ever done this before. And since I'm dealing with (in the test data set) 94 states, I'm basically searching a 94! permutation space. And this is by far my weakest area. So, good times.
Status: lol

1: It's okay if you're not actually asking this.

Comments

[maida-mac.livejournal.com]

Sadly, my knowledge is lacking here.

1 is pretty self-explanatory, though I could argue quite a while about using this sort of system to quantify quality, but that's just me being a photographer. I do find some of their feature set for a quality photo to be very interesting (and true).

2 is pretty clear, but I'm curious as to what the final outcome desired is. Obviously, I know very little about what you've been working on.

3 loses me, 'cause I fail at this type of geekery.

[gfish.livejournal.com]

Certainly I don't think 1 really judges image quality, and I doubt the authors think so either. But it does seem to make a decent stab at it, and more importantly it is looking at exactly the features I'm interested in. We want crisp, well-lit, well-framed pictures for training the SRVC classifier on. Product shots, really. This won't be perfect, but it should do a pretty good job at filtering out some of the dross that comes with any Google Images search.

I'm not sure on the outcome for 2, either. Basically, I just want to get good marks. :) The stuff I've been working on so far is using image features to visually detect and counteract rotation in the platform, much like GYRE. Once I have that working well (it's either very slow or unstable atm) I'll try making it track objects the move in front of it. Since this is all in MATLAB, the interface on something like that becomes tricky.

A Markov process is one where the probability of being in a given state at time t is entirely given by the state it was in at t-1. They're surprisingly useful for linear systems, like language, movement or DNA. Assuming discrete states for convenience, they're defined by a table of starting probabilities for each state, and a table of the transition probability for each state to every other state. How likely is a 'T' to follow a 'Q', etc. So I want to take two sets of tables like that, and assume that the underlying states are the same. (Say, to hint at my real goal, one is taken from the symbols of a syllabic language and one is taken from a transcription of people speaking that language.) I need to find a mapping from states a,c,b in the first set to states 1,2,3 in the second. The permutations add up really quickly, of course. It's very likely untractable unless I come up with something very clever. =\