Paper dealing with an in-depth discussion of bandwidth selection for MS. Here.
Implementing variable dimension bandwidth parameters is proving trickier (and easier) than I had imagined.
I'm setting it up so that the bandwidth parameter, h, for vertex x_i is exactly the maximum edge distance from x_i to its neighbours.
I'm going to then include a parameter that will allow you to scale this edge distance (and thus incorporate a larger sub-graph centred at x_i). This will work proportional to the max edge distance, and conceivable increment by whole numbers.
An issue that I came across is that the adjacent vertices in the mesh topology may not be present in the input points to the MS. This will potentially lead to an under-estimate of the spatial bandwidth for that vertex. The assumption I'm making here is that if a vertex didn't make it into the MS algorithm it's because it wasn't extreme, and we can safely ignore it. This may not be the case if my thresholding is too high, or my spatial extent metric is too noisy.
The bandwidth parameter for the time feature will be completely user-controlled. It remains to be seen whether I should initialise it to some computed "good guess" which can then be changed. The idea here being that the resolution of the clustering along the time axis will depend of the frame density of the animation.
As it is, I've implemented the edgemax parameter stuff. I need to hook them up to the relevant attributes within the Maya node, and set up the time bandwidth and should be good to go.
Bouncing Ball vid
Here is a vid of a ball deforming as it bounces on the ground.
Description:
First (bad) attempt with my clustering algorithm. Issues with false clusters that make no sense, hopefully this will be resolved by defining bandwidth parameters specific to dimension within the mean-shift.