What is the function of the grid points? Can it be used to subsample a point cloud in a voxel downsample way?

[liu]

Hi,

Thank you for the great work!

Currently I need to find the k nearest neighbor for a point cloud, then I need to downsample the point cloud in a voxel downsample way, can the grid output be use for the downsampling?

[Lixin Xue]

Hi,

I hope I understand your question correctly. Here is the documentation for the returned grid structure:

FRNN/frnn/frnn.py

Lines 217 to 223 in 59a4c8f

       grid: A namedtuple of 4 tensors for reusing the grid if necessary/possible. 
             Users don't need to know what's inside. 
             _GRID = namedtuple("GRID", "sorted_points2 pc2_grid_off sorted_points2_idxs grid_params") 
               sorted_points2 is a spatially sorted version of points2; 
               sorted_points2_idxs records the mapping between points2 and sorted_points2; 
               grid_params characterizes the constructed grid; 
               pc2_grid_off records the start and the end indices of sorted_points2 for each grid 

In sorted_points2, points are sorted according to their voxel index. Therefore, with pc2_grid_off you will know the points inside one specific voxel (Here 2 are hard-coded for 2D points, you might want to change to 3):

FRNN/frnn/csrc/grid/find_nbrs.cu

Lines 50 to 60 in 3e370d8

 int p2_start = pc2_grid_off[n * G + cell_idx]; 
 int p2_end; 
 if (cell_idx + 1 == grid_total) { 
   p2_end = lengths2[n]; 
 } else { 
   p2_end = pc2_grid_off[n * G + cell_idx + 1]; 
 } 
 for (int p2 = p2_start; p2 < p2_end; ++p2) { 
   float2 diff; 
   diff.x = points2[n * P2 * 2 + p2 * 2] - cur_point.x; 
   diff.y = points2[n * P2 * 2 + p2 * 2 + 1] - cur_point.y; 

And then you can do whatever you want to downsample the points inside each voxel.

[liu]

Thanks a lot! Are we allowed to change the initial grid size before fitting the knn?

[Lixin Xue]

Yes, the search radius r and radius_cell_ratio jointly determine the grid size:

FRNN/frnn/frnn.py

Line 77 in 3e370d8

 cell_size = r[i].item() / radius_cell_ratio 

For KNN you should find a radius large enough so that you can find K neighbors within the radius for sure. When the radius_cell_ratio is an integer, the method runs more efficiently since it search 2*ceil(radius_cell_ratio)+1 grids on one dimension.