Try using DBSCAN for initial clustering

[kbarbary]

In the multi-ellipsoidal clustering, clusters are created by recursively splitting points into two clusters. In some (contrived) examples, this results in strange partitioning. Example:

In this case, it would clearly be better to bound the center points in a single ellipsoid. What probably happened here is that the points were initially split into two clusters vertically, meaning that the center mode was split into two right away and so could never be reassembled.

It might be helpful to do initial clustering using an algorithm with a non-fixed number of clusters such as DBSCAN. However DBSCAN would not split curving degeneracies into multiple clusters, so we'd still need to use the current algorithm for that.

I'm not sure how big a problem this is in "real-world" problems.