CS-544 Optimization in Computer Vision

1. Warmup: Build an implementation of grab-cut (paper on the website). You may use any flow-cut package you like, or you could build your own. The main point of this exercise is to get used to the image costs, etc. You may find it easier to use a tri-map, but you do not need to build an interactive interface to make tri-maps. You could use GIMP or something.
2. Now extend this to cover two, distinct, foreground objects. There will be three classes - obj1, obj2, background - so you can't use exact inference and your tri-map becomes a quad-map. Use alpha-expansion with your flow-cut package to get an approximate solution
3. For any of your images, is your solution the true optimum? how would you know?
4. Harder optional questions:
   • Can you use branch-and-bound to improve your solution?

The following paper has a neat trick to improve the segment boundaries; you adjust the quadratic term to account for contrast (p432, eqn 12).

• Interactive Image Segmentation Using an Adaptive GMMRF Model Andrew Blake, Carsten Rother, M. Brown, Patrick Perez and Philip Torr, ECCV 2004

This is a broad MP intended to be educational. I will grade on quality of experimental concept and of argument from data. Note change of deadline Submit PDF's by 23h59 Sun May 13. Email them to Xiaoming at xz23@illinois.edu, with "CS544" in the header