Non-rigid cpd with bunny point clouds about probreg

[neka-nat]

I tried it in my environment and it seemed to be fine.
Is there any difference?

import numpy as np
import open3d as o3
import transformations as trans
from probreg import cpd
from probreg import callbacks
import utils

source, target = utils.prepare_source_and_target_nonrigid_3d('bunny-x.txt',
                                                             'bunny-y.txt', 0.1)

cbs = [callbacks.Open3dVisualizerCallback(source, target)]
tf_param = cpd.registration_cpd(source, target, 'nonrigid',
                                lmd=1.0e-9, maxiter=100,
                                tol=0.000000000000001, callbacks=cbs)

[sandwich25]

For me, I used bunny-y as source and bunny-x as target. Can you try that please?

[neka-nat]

Thank you!
I reproduced.
And I tuned the parameters again.
Try it out.

import numpy as np
import open3d as o3
import transformations as trans
from probreg import cpd
from probreg import callbacks
import utils

source, target = utils.prepare_source_and_target_nonrigid_3d('bunny-y.txt',
                                                             'bunny-x.txt', 0.1)

cbs = [callbacks.Open3dVisualizerCallback(source, target)]
tf_param = cpd.registration_cpd(source, target, 'nonrigid',
                                lmd=1.0e-11, maxiter=100,
                                tol=0.000000000000001, callbacks=cbs)

[sandwich25]

There are still two points not aligned.

[neka-nat]

It's not easy to adjust.
What about the following?

tf_param = cpd.registration_cpd(source, target, 'nonrigid',
                                lmd=1.0e-15, beta=20.0, maxiter=100,
                                tol=0.000000000000001, callbacks=cbs)

[sandwich25]

Two dots still not aligned. It's okay if you don't have the solution. Any idea on the reason this is happening?

[neka-nat]

It's a very divergent problem with CPD because the transformation is so large.
BCPD allows for a more stable convergence.

[sandwich25]

Thanks!