|
Yay for 2 conference submissions! |
8/4/05 11:07am
|
So in the last week and a half Brian and I were able to pound out a 500 words abstract based on the implementation and inner-working
of our seeded region growing algorithm. We sent that one to the SPIE: Electronic Imaging conference that'll be in San Jose sometime
in January of next year. Hopefully we get accepted because it'd be really exciting to present this stuff we've been working on and
see what others have to say about it. We also just got in anothe submission to the SPIE: Medical Imaging conference with a paper
focusing on minimal user intervention and one-click segmentation. So that one is more geared towards the medical folk and how our
algorithm performed on regions like the kidneys, liver, and spleen with only a single click of the user. Hopefully we get accepted
to that conference too. It would be in February in San Diego. Good times...
-Nathan
|
|
Things are getting interesting! |
7/15/05 11:44am
|
Interesting stuff has been brewing in the research lab recently. Brian and I have tried using the Euclidean Distance,
the Jeffrey Divergance, the Mahalanobis Distance, and now we're onto something a wee bit different for a distance metric. We've
started to examine the distribution of our texture descriptor data in order to find a better distance metric and we have come up with
a metric that essentially takes the probability between two points in a distribution. This has been doing wonders
for our results. It's helping us onto something interesting for stopping criteria. We have been looking at real-time
histograms of these distances as we have been able to identify cut-offs that represent leaking points. -Nathan
|
|
Implementation |
6/24/05 10:58pm
|
So we realized that our data wasn't very much up to par. We were thinking about doing classification and the
pre segmented images were really not segmented well at all. We figured it would be pretty cool to do our own
form of segmentation and the idea of region growing came to mind. We figured we could compare neighboring pixels
with an all pair pixel matching method to form our cooccurrence matrix from a 5x5x5 cube around each individual
pixel in order to calculate pixel level descriptors. We iterated through in a selection sort style manner in order
to proveide a sense of direction and distance to our pairing algorithm that wouldn't return to same results for
images such as a checkerboard and half white half black image.
We started implementing, and were able to calculate cooccurrence matrices and texture descriptors for each pixel
as well as create a breadth first search in order to do our region growing in a volumetric sense. We've done quite
a bit of work and both Brian and I are usually pretty worn out by the end of the day. I'm really pleased by the progress
we've made so far. Now we just need to find out how to normalize our texture descriptors so that we can compare
neighboring pixels and then do some sort of clipped binning, instead of the linear binning we are doing now, which
shouldn't be too difficult. -Nathan
|
|
First Post! |
6/24/05 10:02am
|
Well, it turns out that I'm on the volumetric segmentation/classification side of things with Dr. Furst,
and it turns out that it's pretty cool stuff. I'm working with Brian on it and we've seemed to have gotten
a lot accomplished in this first week. We were able to dig through code previously used on the project
and recreate old results and get a feel for where the previous folk were taking things.-Nathan
|
|
|