The first step, making a 3-D correlation Matrix, is complete. It works by sliding one image past the other and doing a correlation each time. The range to slide the image needs to be determined, well, by hand. I guess the idea is that for a given set, you had an idea of the orientation and how far apart the pictures were taken. That's working fine for us now.
The next part is difficult. We'll try to use edge info and multiscale info with this correlation matrix to somehow interpret the images. This involves finding the best path through the correlation cube along each epipolar line. We'll see. After that, some sort of median smoothing/interpolation is done to even it out. We'll probably use a wavelet reduction (just for the heck of it) because then we can get out multiscale correlation matrices, as well as a bunch of edge info.
Unfortunately, we cannot exploit all the speed built into the algorithms. It uses nested loop iterations to do several operations at once, whereas in matlab, convolution and matrix multiplies are faster using built-in methods in separate iterations.
Next week, we'll have it identifying yo ugly ass, watch out!
Peace and Filtering!