Once the code has been modified so that M and E steps are merged and covariance matrices don't have to be stored, we can use much larger numbers of clusters. In particular, we could consider one initial cluster for each unique binary mask (or maybe we should reduce them a little bit). This could potentially solve a lot of the problems with big cluster masks.
Once the code has been modified so that M and E steps are merged and covariance matrices don't have to be stored, we can use much larger numbers of clusters. In particular, we could consider one initial cluster for each unique binary mask (or maybe we should reduce them a little bit). This could potentially solve a lot of the problems with big cluster masks.