Project __ Writeup
Eli Bosworth (eboswort)
February 26, 2010
My Graph Cut created a matrix A of size (image height*image width) by (image height*image width) where every entry represented
the cost of making that pixel into the border between the source and target images. The weight for an edge was decided by the sum of the differences
between the source and target images at each of the two pixels. That matrix was passed to the maxflow algorithm, along
with two vectors representing the areas of the image that must come from the source and target respectively. Maxflow returned a mask
that represented the lowest cost border of the source image. I then combined the source and target image using that mask and the fast
poisson provided by the TA's.
- I made a small addition to the algorithm where instead of just setting the weight as I described
above, I ran a sobel filter over the source image and for each pixel subtracted 1/10th the value returned from the sobel filter from the weight of the edges
connected to it. My intention was to influence max flow to cut the source image along it's edges. I was hoping for a close cut around the central
object in the source but ended up with strange masks. In some cases, it helped (The horse and the eiffel tower). Had it not been Friday night I would've liked
to experiment a little more. I tried also factoring in target edges but that was awful for all the images.
Results Images:
The higher of each image pair was done with my edge detecting add-on the one below was done without. The masks are on the right
