Using Graph Cut algorithm

Question

I want to use Graph Cut algorithm to label some Video Cubids. Here is my problem.

As far as I understand, GraphCut has an energy function being called D (cost per element), and a V pairwise that is cost of two adjacent naighbours. Now, I have processed video and I have some idea to what to set as V (like euclidean distance) , but I have no idea what I could use for energy function D. It should be noted, I don't have train data, so I am not be able to compare input data to a model or foreknown distribution to obtain D.

Answer 1

The short answer is that you need D(l,s) to be related to how likely it is for the site s to have the label l. How to model that and what are acceptable things to use for D spans many research areas. For example there are many papers that study reasonable D's for stereo, segmentation, denoising, etc. A good reference to understand these issues would be the original Veksler, Boykov, Zabih paper. Once you understand this basic paper you can read more sophisticated works related to how to user Graphcuts applied to the particular problem you want to solve.