Thursday, July 17, 2008 1:15 PM
IPCV 08' Part 2 - Novel Approaches
I've managed to rig up Steve's cell to my laptop and I once again have Internet Access.
From the time of my last post, most of the talks have been about applications of machine learning. This usually involves a neural network, PCA (Principle Component Analysis), wavelets and/or some novel feature or metric. Some presentations that stood out were:
A very large project for the detection of pollution outside of Las Vegas. This is done by flying an unmanned aircraft over the area and using a specifically designed system which works by breaking light up into subspectrums via prisms. The pollution can then be detected by the intensity of these subspectrums.
A circularity measure for digital curves. In this paper the author developed a fast two-color approach for measuring how round a document component is based on how much of an object can be completely contained in a circle.
Using minimal spanning trees to maximize the compression of a group of similar images. I found this interesting mainly because it reminded me somewhat of the PNG preprocessing which I implemented in our PDF stuff. The distance on the MST is used to determine which group the images belong to. These groups then have average images calculated for them which allows cheap-to-compress distance images to be stored.
Use of a Discrete Tchebichef Transform (Instead of a Discrete Cosine Transform) in MPEG2 compression. This was mostly interesting because, while the Tchebichef transform performed similarly on most images (in terms of compression) it performed much better on unnatural images. The big disadvantage was that DTT is not as widely used and so does not have algorithms which are as highly optimized as DCT.
Steve's talk is still to come. Because his talk is the last of the conference, he came up with a scheme to increase viewership by offering free candy. Only time will tell how this works out.