2
$\begingroup$

I try to use graph-cut method but I got some trouble with it. I want to segment an image into foreground and background. For each pixel $x$, I got the probability that it belongs to the background $p(x|\phi_b)$ and to the foreground $p(x|\phi_f)$ by kernel density estimation (KDE). Here is the log-posterior :

$$ \log p(\mathcal{L}|x) = \left( \sum^N_{i=1} \sum^N_{j=1} \lambda(l_il_j + (1-l_i)(1-l_j)) \right) + \sum^N_{i=1} \log\left( \frac{p(x_i|\phi_f)}{p(x_i|\phi_b)} \right) l_i $$

where $\mathcal{L} = [l_1...L_N]$, $N$ is the number of pixel in the image.

If I understand, the first term is the pairwise cost, for smoothness and determine the weight on edges between a pixel and its neighbors. According to the equation, I put lambda as a weight on each edges.

The second term is the unary cost and determine the weight on edges between a pixel and the source and the sink. Here, the weight between a pixel and the source is the same that between the pixel and the sink ? Am I right ?

Thanks

$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Browse other questions tagged or ask your own question.