Random seed works at 0.15.1 segment_plane but not works at 0.16.0

[mcmingchang]

Checklist

• I have searched for similar issues.
• For Python issues, I have tested with the latest development wheel.
• I have checked the release documentation and the latest documentation (for master branch).

Describe the issue

Random seed works at 0.15.1 segment_plane but not works at 0.16.0

Steps to reproduce the bug

#### in 0.15.1
import numpy as np
import open3d as o3d
pcd = o3d.io.read_point_cloud('1.ply')
pcd_down = pcd.voxel_down_sample(voxel_size=0.01)
plane_model, inliers = pcd_down.segment_plane(distance_threshold=0.02,
                                              ransac_n=5,
                                              num_iterations=5000, seed=1)
plane_model, inliers2 = pcd_down.segment_plane(distance_threshold=0.02,
                                              ransac_n=5,
                                              num_iterations=5000, seed=1)
print(np.all(np.array(inliers) == np.array(inliers2)))


##### in 0.16.0
import numpy as np
import open3d as o3d
o3d.utility.random.seed(1)
pcd = o3d.io.read_point_cloud('1.ply')
pcd_down = pcd.voxel_down_sample(voxel_size=0.01)
plane_model, inliers = pcd_down.segment_plane(distance_threshold=0.02,
                                              ransac_n=5,
                                              num_iterations=5000)
plane_model, inliers2 = pcd_down.segment_plane(distance_threshold=0.02,
                                              ransac_n=5,
                                              num_iterations=5000)
print(np.all(np.array(inliers) == np.array(inliers2)))

Error message

None

Expected behavior

The result of 0.16 should be the same as that of 0.15

Open3D, Python and System information

None

Additional information

Random seed works at 0.15.1 segment_plane but not works at 0.16.0

[yuecideng]

Since 0.16.0, the segment_plane function is parallel and the iteration is changed with the estimated inliers ratio, which may lead to inconsistent results under the same random seed.

[mcmingchang]

You mean it's not a bug？But setting seed is to achieve a stable result.

[yuecideng]

You may try to set the new arg probability=1, which will force the iteration to be executed equals to num_iterations. Then the results will be consistent with the seed you set.

[mcmingchang]

After experiments, the results are different each time.

[yuecideng]

The random seed is singleton so you should reset the random seed before segment_plane at the second time.

import numpy as np
import open3d as o3d
o3d.utility.random.seed(1)
pcd = o3d.io.read_point_cloud('1.ply')
pcd_down = pcd.voxel_down_sample(voxel_size=0.01)
plane_model, inliers = pcd_down.segment_plane(distance_threshold=0.02,
                                              ransac_n=5,
                                              num_iterations=5000,
                                              probability=1)
o3d.utility.random.seed(1)
plane_model, inliers2 = pcd_down.segment_plane(distance_threshold=0.02,
                                              ransac_n=5,
                                              num_iterations=5000,
                                              probability=1)
print(np.all(np.array(inliers) == np.array(inliers2)))

[mcmingchang]

thank you

[mcmingchang]

For large plane fitting, it is not work at 0.16

[conorsim]

I have also observed this issue with 0.16 and 0.17

[TheExDeus]

I set seed before each call and probability=1.0 - and it still doesn't make this deterministic. I also get different value every time (and in some cases it can vary a lot if the points are sparse). This means I cannot make tests. Any idea on how to fix this?

This means I might need to revert to 0.15, but that doesn't seem like a long term solution.

[yuecideng]

Hi @TheExDeus @conorsim, could you provide the data or the code snippet to reproduce the bug?

[conorsim]

@yuecideng if I test your code above with version 0.17.0 I don't get the same values. I made a couple of changes:

1. To initialize a point cloud of random points since I don't have a .ply file handy
2. To compare plane_model since inliers doesn't always have the same shape (due to the this issue)

import numpy as np
import open3d as o3d

pcd = o3d.geometry.PointCloud()
pcd.points = o3d.utility.Vector3dVector(np.random.uniform(1,10,(1000,3)))

for _ in range(10):
    o3d.utility.random.seed(1)
    pcd_down = pcd.voxel_down_sample(voxel_size=0.01)
    plane_model, inliers = pcd_down.segment_plane(distance_threshold=0.02,
                                                  ransac_n=5,
                                                  num_iterations=5000,
                                                  probability=1)
    o3d.utility.random.seed(1)
    plane_model2, inliers2 = pcd_down.segment_plane(distance_threshold=0.02,
                                                  ransac_n=5,
                                                  num_iterations=5000,
                                                  probability=1)
    print(np.all(np.array(plane_model) == np.array(plane_model2)))

When I run this, the output I got was

False
False
False
False
True
False
False
False
False
False

So, in this run it was correct only 1/10 times.

[yuecideng]

Hi @conorsim, you are right! this may be caused by the update of best estimated model among multi threading. I will try to investigate it.

[yuecideng]

The issue has been solved by PR #6308.