I have a general idea of Gaussian Mixture Models. My understanding:
GMM is a way of clustering data points which, unlike K means clustering, soft assigns them under different distributions by calculating posterior $P(\theta_k|X)$ where $\theta_k$ are the parameter of distribution k. This is done by Expectation Maximization where the log likelihood function is maximized with respect to parameter of each distribution and updated till the algorithm converges.
I still don't understand how the "user assisted" part of segmentation works. Also, it was mentioned in my lectures that each class has its own Mixture of Gaussians that is used to calculate posterior probabilities and the pixel is assigned to to class with higher probability but I can't wrap my head around this statement.
Thanks.
