Implementation of 2 downsampling methods for Point Cloud - random_downsample; voxel_filter

[YuqingLi-hub]

The downsampling method is used to unify data size and reduce computational cost. Here, we implement two commonly used methods.

Random_downsample:
This function takes two arguments: a point cloud and the desired number of points after downsampling. It outputs a randomly selected downsampled point cloud.

Voxel_filter:
This function conducts voxel filtering on a provided point cloud. It has three arguments: the given point cloud, voxel size in three dimensions, and an optional boolean value that indicates whether to randomly select a point from each voxel as a representative or use the mean of all points in each voxel. The output is the downsampled point cloud, where each point serves as the representative point in its corresponding voxel.

[wang-chen]

I can see the status of this PR is still in progress, if you are ready for review, please let me know.

[YuqingLi-hub]

I can see the status of this PR is still in progress, if you are ready for review, please let me know.

Hi Wang Chen,
It's ready to review now.

[wang-chen]

Hi, I renamed some names and tried to make these functions work for arbitrary dimensional space such 2D.
You also need to add auto tests for those functions, such as this test.

[wang-chen]

Hi @YuqingLi-hub , Thank you so much for this update. The provided testing codes are to be working fine.
I tried the following test codes, it seems that they cannot run for batched tensors.

points = torch.randn(2, 10, 3)
num = 5
result = pp.geometry.random_filter(points, num)
result = pp.geometry.voxel_filter(points, [1., 1., 1.])

For random_filter, I changed to the following code, it seems working.

def random_filter(points, num):
    assert points.size(-1) >= 1, "The last dim of the points should not less than 1."
    assert num <= points.size(-2), "Number of points to sample must not larger than " \
        "the number of input points."
    
    indices = torch.randperm(points.size(-2))[:num]
    return points[..., indices, :]

I am not sure how to change voxel_filter, as we may need to change multiple places to make it to be compatible with batched operations.

[wang-chen]

Until now, I have tried:
Change this line to

minp = torch.min(points[..., :dim], dim=-2, keepdim=True).values

this line to

voxel_counts = torch.zeros(unique_indices.shape[:-1], dtype=points.dtype)

You may want to access a specific dimension with negative numbers, as we may have multiple batches, e.g., points can be a shape of [... ,5,3], where ... can be any dimensions.