This type of algorithm could be useful for research in situations where people need good segmentation to accomplish other tasks such as object regonition or modeling. The results are obviously better than simple background subtraction or color-based segmentations and could be used to generate "good" data. However, in practice, especially in their video example, this sort of segmentation seems impractical.

Their gestalt example indicates that the behavior of this method can depend heavily on the value for lambda (the region term). Would the user set this value?

How would the graph cut work for a 3x3 image that had a horizontal "bar" in the center:

0 0 0
1 1 1
0 0 0

where 0 is background and 1 is object. If you have time tomorrow, could you walk through this example?

Edited by author 10-21-2004 12:15 PM
It seems to me that the hard constraints can be overruled, as they are not assigned an infinite cost. But if this is the case, the user assigned bits are not believed to be 100% true. It is interesting that the algorithm doesn't trust the user.

Also, why separating two objects from each other is not an issue but we are only interested in separation of background from objects (I think because it is more difficult)?

Edited by author 10-21-2004 01:53 PM
What is meant by "Isolated Connected Blobs" ?
(fourth page first paragraph)

The histogram for "object" and "background" intensities seems an important part of the finding values for the boundary and region penalites (page 5).

I'm curious whether only a rough estimate is required (dark object on light background, vs light object on dark background), and how sensitive these penalty terms are to choices of seed pixels that are representative of the whole object.

jkjgtjhgtrjjrhgjhtrj and more info
D