Question about initialization of nonrigid transformation

[Buglakova]

Hello,

Thank you for a great library!

I'm using CPD to do nonrigid registration of 3D point clouds. First I register them rigidly, so they match pretty well at the start. According to my benchmark, the best parameters to be used for my datasets are like this:

cpd.registration_cpd(source_pt, target_pt, use_cuda=use_cuda, maxiter=args.maxiter, tf_type_name="nonrigid", w=0.0001, beta=1000, lmd=0.1)

When I look at the output of each iteration, in the beginning the moving point cloud gets distorted to fit completely inside the fixed point cloud. The resulting registration looks good but I wonder what happens before the first iteration and if I might be using the algorithm wrong.

Here is an illustration of what I mean. It's all points of the point clouds projected on (x, y) plane. On the second plot pink and blue colors should be the other way around but I hope you see what I mean.