How to set the distance threshold in registration about probreg HOT 6 CLOSED

Hi @LogWell ,

Thank you for reporting.
Which method do you use? (CPD, GMMReg, SVR, GMMTree or Filterreg)

from probreg.

left: probreg

tf_param, _, _ = cpd.registration_cpd(source, target, 'rigid', callbacks=cbs)

right: open3d, with threshold=0.02

data: please change .txt to .ply
0_0_00 (copy).txt
bch0002hd2o01p01s__result2__s (copy).txt

registration with scale is very useful

from probreg.

Thank you for the data.
I noticed that this data are deformable objects, so I tried affine transformation registration.
It's a little better, I think, isn't it?

import copy
import numpy as np
import open3d as o3
from probreg import cpd

source = o3.io.read_point_cloud('data1.ply')
target = o3.io.read_point_cloud('data2.ply')
source = source.voxel_down_sample(voxel_size=0.02)
target = target.voxel_down_sample(voxel_size=0.02)
result = copy.deepcopy(source)

tf_param, _, _ = cpd.registration_cpd(source, target, 'affine')
result.points = tf_param.transform(result.points)

target.paint_uniform_color([0, 1, 0])
result.paint_uniform_color([0, 0, 1])
o3.visualization.draw_geometries([target, result])

from probreg.

If a complete mesh is registered with an incomplete point cloud, a distance threshold is still needed (to eliminate some incorrect corresponding points), right?

For the mesh of human body, if we want to compare the results of different algorithms, we need to register first. Generally, we only need to rotate, scale and translate. The affine transformation will make the comparison unfair.

from probreg.

As you say, in ICP you can set a distance threshold on the distance between the source point and the nearest neighbor point of the target point, but in CPD it is difficult because it expresses the distance between points with probability.
Take a look at the paper if you like.
The paper also lists registration results with incomplete data (Fig.6), so isn't the distance threshold the only thing that matters?

If you want to make a fair comparison, it may be better to use rigid deformation data.
Because when shown the above comparison, you can't categorically say that the right one is better.
The result on the left is certainly a bit of a leaner body part, but it looks more like it is trying to align their left foot.

from probreg.

I see. Thank you very much~

from probreg.