hey !
does anyone has experience with the Restricted Boltzmann Machine component of the Lunchbox plugin ?
I would like to use it to find corresponding points in two different point clouds:
a cloud that lies flat in XY-plane and a corresponding cloud after a non-linear transformation.
the tranformed cloud is the result of a 3d-scan, so the points in both lists do not have the same order, otherwise this would be a no brainer…

any tips/explanation for the Momentum, Learning Rate, Decay and Alpha input ?

I cannot share my actual data, but I simplified it to give you an idea of it:

I have a starting point cloud coming from nodes of a cell structure/mesh (blue). This comes from a workpiece which gets physically deformed. So after 3D-scanning its final shape, I get the second point cloud (red) that’s somehow orientated in space.

The order of the points in both lists are - of course - not identical !

Now I want to find the corresponding points of the blue set of points, to recreate the cell structure and evaulate the deformation of each edge. And in my case I have to work with over 700 points.

We found some interesting articles about this process as part of animation projects, so that’s how the Restricted Boltzmann Machine came up.
Unfortunately, there is no helpful documentation of this component provided by Lunchbox. At least I couldn’t find anything helpful yet…

I’ve done something similar with Kangaroo. In my case, the mesh had four corners and unsorted vertices. All vertices get pulled from their 3D position to the XY plane and the edges are equalized. The resulting points are in a nice grid and can be sorted.

The mesh was from a file someone else uploaded here in the forum. I don’t know how it was created. Certainly something went wrong in the process.

Your ‘pringle’ shape is somewhat planar and nearly circular. The example below works fine but if points are closer together it will be more difficult to separate them and find pairs of identical points.