Add uniform and random down sample methods in Tensor PointCloud #5202

yuecideng · 2022-06-15T14:34:23Z

Update
Use tensor slicing stead of indexing for uniform sampling, which makes it more faster.

This change is

update-docs · 2022-06-15T14:34:27Z

Thanks for submitting this pull request! The maintainers of this repository would appreciate if you could update the CHANGELOG.md based on your changes.

yuecideng · 2022-06-16T07:12:33Z

Benchmark for UniformDownSample.

ssheorey · 2022-06-16T14:42:13Z

cpp/open3d/t/geometry/PointCloud.cpp

+    return pcd_down;
+}
+
+PointCloud PointCloud::RandomDownSample(double sampling_ratio) const {


For reproducible experiments, we need the ability to set the random seed for all RNG calls by Open3D. One way to achieve this is with a singleton RNG, with some care for concurrent access.

CUDA has an RNG that will likely be faster for CUDA tensors.

reyanshsolis · 2022-06-17T09:21:17Z

I appreciate the effort, but I was wondering if there is any use case for this functionality?

Here uniform is misleading, as the point cloud is unorganized data, so removing every k-th point, will not downsample the point cloud uniformly.

Do we have a reference implementation/inspiration of such functionality in PCL or other 3D data processing packages?

reyanshsolis · 2022-06-17T09:27:08Z

Voxel down-sampling is actually a uniform down-sample in a way.
However, it lacks the option of averaging or sampling, i.e. control if the new point attributes are the mean of the points and their attributes within a voxel, or it's one of the existing points (sampled).

I guess users might appreciate an improved voxel-down-sample functionality over other filtering options.

yuecideng · 2022-06-17T09:45:35Z

The uniform here means taking a subset of point cloud with the same interval w.r.t original order. The use case is to provide a simple interface for user to fast down sample a very large pcd (eg. 10M points, particularly for industrial scenario, this can be commomly seen). Sometimes it is no worth to use VoxelDownSample when consider time comsuming. Beside, we also provide this function in legacy version.

However, it lacks the option of averaging or sampling, i.e. control if the new point attributes are the mean of the points and their attributes within a voxel, or it's one of the existing points (sampled).

I agree with this because currenly the voxel down sample method in tensor based pcd only return coordinates of voxels, leading to the inconsistent with legacy version and other library's behaviour.

reyanshsolis · 2022-06-19T16:15:07Z

cpp/open3d/t/geometry/PointCloud.h

    /// \param voxel_size Voxel size. A positive number.
    PointCloud VoxelDownSample(double voxel_size,
                               const core::HashBackendType &backend =
                                       core::HashBackendType::Default) const;

+    /// \brief Downsamples a point cloud uniformly


please elaborate on the description such as:
Downsamples a point cloud, by selecting every kth index point and its attributes.

Similarly for RandomDownSample

reyanshsolis · 2022-06-19T16:18:20Z

cpp/pybind/t/geometry/pointcloud.cpp

@@ -198,6 +198,18 @@ The attributes of the point cloud have different levels::
            },
            "Downsamples a point cloud with a specified voxel size.",
            "voxel_size"_a);
+    pointcloud.def(
+            "uniform_down_sample",


You may use the following:
Instead of

pointcloud.def( "uniform_down_sample", [](const PointCloud& pointcloud, const size_t every_k_points) { return pointcloud.UniformDownSample(every_k_points); }, "Downsamples a point cloud uniformly.", "every_k_points"_a);

you may try:

pointcloud.def( "uniform_down_sample", &UniformDownSample, "Downsamples a point cloud uniformly.", "every_k_points"_a);

similarly to other functions.

Also, test the same if it is working, you may find similar use in pybind folder for reference.

reyanshsolis

Reviewable status: 0 of 5 files reviewed, 5 unresolved discussions (waiting on @yuecideng and @yxlao)

cpp/tests/t/geometry/PointCloud.cpp line 770 at r2 (raw file):

    core::Device device = GetParam();

    // Value test

Add a .

cpp/tests/t/geometry/PointCloud.cpp line 789 at r2 (raw file):

    core::Device device = GetParam();

    // Value test

same

yuecideng

Reviewable status: 0 of 5 files reviewed, 5 unresolved discussions (waiting on @reyanshsolis, @yuecideng, and @yxlao)

cpp/tests/t/geometry/PointCloud.cpp line 770 at r2 (raw file):

Previously, reyanshsolis (Rishabh Singh) wrote…

Add a .

Done

cpp/tests/t/geometry/PointCloud.cpp line 789 at r2 (raw file):

Previously, reyanshsolis (Rishabh Singh) wrote…

same

Done

heethesh · 2022-06-22T00:50:08Z

#4849

yuecideng · 2022-06-25T14:22:42Z

Update with using open3d random singleton. CUDA implementation will be added in the future PR.

yxlao

Reviewed 1 of 2 files at r2, 2 of 3 files at r4, 1 of 1 files at r5, all commit messages.
Reviewable status: 4 of 5 files reviewed, 8 unresolved discussions (waiting on @reyanshsolis, @ssheorey, @yuecideng, and @yxlao)

cpp/benchmarks/t/geometry/PointCloud.cpp line 123 at r5 (raw file):

    pcd = pcd.To(device);

    // Warp up

typo: warm up.

cpp/pybind/t/geometry/pointcloud.cpp line 204 at r5 (raw file):

                   "Downsamples a point cloud by selecting every kth index "
                   "point and its attributes.",
                   "every_k_points"_a);

Same as below, call docstring::ClassMethodDocInject

cpp/pybind/t/geometry/pointcloud.cpp line 208 at r5 (raw file):

                   "Downsample a pointcloud by selecting random index point "
                   "and its attributes.",
                   "sampling_ratio"_a);

We need to call docstring::ClassMethodDocInject to enable the docs. In our case, we need to document sampling_ratio by passing in a dictionary.

Code quote:

docstring::ClassMethodDocInject

yuecideng

Reviewable status: 4 of 5 files reviewed, 8 unresolved discussions (waiting on @reyanshsolis, @ssheorey, @yuecideng, and @yxlao)

cpp/pybind/t/geometry/pointcloud.cpp line 208 at r5 (raw file):

Previously, yxlao (Yixing Lao) wrote…

We need to call docstring::ClassMethodDocInject to enable the docs. In our case, we need to document sampling_ratio by passing in a dictionary.

Should we also add docstring::ClassMethodDocInject to the recently added functions, like SelectByMask, SelectByIndex, RemoveRadiusOutliers? I think it is better to add them all.

yxlao

Reviewable status: 3 of 5 files reviewed, 9 unresolved discussions (waiting on @reyanshsolis, @ssheorey, @yuecideng, and @yxlao)

cpp/open3d/t/geometry/PointCloud.cpp line 359 at r6 (raw file):

                false, false);
    } else {
        utility::LogError("RandomDownSample is not supported on GPU.");

LGTM, after fixing this. We can copy the indices from CPU to CUDA for now. This is required to pass the CUDA CI test.

yxlao

Reviewable status: 3 of 5 files reviewed, 9 unresolved discussions (waiting on @reyanshsolis, @ssheorey, @yuecideng, and @yxlao)

add uniform and random down sample methods

63390cf

yuecideng added 2 commits June 16, 2022 13:34

improve uniform sampling performance

885bb54

add benchmark for uniform down sample

710a8a7

yuecideng requested review from reyanshsolis and yxlao June 16, 2022 07:13

ssheorey reviewed Jun 16, 2022

View reviewed changes

reyanshsolis reviewed Jun 19, 2022

View reviewed changes

reyanshsolis requested changes Jun 19, 2022

View reviewed changes

apply changes

e9169ee

yuecideng commented Jun 20, 2022

View reviewed changes

heethesh mentioned this pull request Jun 22, 2022

Add Doppler ICP in tensor registration pipeline #5237

Merged

4 tasks

yuecideng added 2 commits June 25, 2022 17:57

Merge branch 'master' into yueci/tpcd-sampling-method

a6495ce

use open3d random singleton in random sampling

d592b36

apply style

929a374

yxlao requested changes Jun 27, 2022

View reviewed changes

yuecideng commented Jun 27, 2022

View reviewed changes

update pybind docs

67c315a

yxlao requested changes Jun 28, 2022

View reviewed changes

minor change

ed99052

yxlao approved these changes Jun 28, 2022

View reviewed changes

yxlao merged commit 3bf5bf0 into master Jun 28, 2022

yxlao deleted the yueci/tpcd-sampling-method branch June 28, 2022 06:02

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Add uniform and random down sample methods in Tensor PointCloud #5202

Add uniform and random down sample methods in Tensor PointCloud #5202

yuecideng commented Jun 15, 2022 •

edited

Loading

update-docs bot commented Jun 15, 2022

yuecideng commented Jun 16, 2022 •

edited

Loading

ssheorey Jun 16, 2022

reyanshsolis commented Jun 17, 2022 •

edited

Loading

reyanshsolis commented Jun 17, 2022

yuecideng commented Jun 17, 2022

reyanshsolis Jun 19, 2022

yuecideng Jun 20, 2022

reyanshsolis Jun 19, 2022

yuecideng Jun 20, 2022 •

edited

Loading

reyanshsolis left a comment

yuecideng left a comment

heethesh commented Jun 22, 2022

yuecideng commented Jun 25, 2022

yxlao left a comment

yuecideng left a comment

yxlao left a comment

yxlao left a comment

Add uniform and random down sample methods in Tensor PointCloud #5202

Add uniform and random down sample methods in Tensor PointCloud #5202

Conversation

yuecideng commented Jun 15, 2022 • edited Loading

update-docs bot commented Jun 15, 2022

yuecideng commented Jun 16, 2022 • edited Loading

ssheorey Jun 16, 2022

Choose a reason for hiding this comment

reyanshsolis commented Jun 17, 2022 • edited Loading

reyanshsolis commented Jun 17, 2022

yuecideng commented Jun 17, 2022

reyanshsolis Jun 19, 2022

Choose a reason for hiding this comment

yuecideng Jun 20, 2022

Choose a reason for hiding this comment

reyanshsolis Jun 19, 2022

Choose a reason for hiding this comment

yuecideng Jun 20, 2022 • edited Loading

Choose a reason for hiding this comment

reyanshsolis left a comment

Choose a reason for hiding this comment

yuecideng left a comment

Choose a reason for hiding this comment

heethesh commented Jun 22, 2022

yuecideng commented Jun 25, 2022

yxlao left a comment

Choose a reason for hiding this comment

yuecideng left a comment

Choose a reason for hiding this comment

yxlao left a comment

Choose a reason for hiding this comment

yxlao left a comment

Choose a reason for hiding this comment

yuecideng commented Jun 15, 2022 •

edited

Loading

yuecideng commented Jun 16, 2022 •

edited

Loading

reyanshsolis commented Jun 17, 2022 •

edited

Loading

yuecideng Jun 20, 2022 •

edited

Loading