Upsampling / Downsampling pcd files while keeping intensity field

[destoswa]

Hello,
I am having an issue. As said in the title, I am trying to upsample or downsample pcd files to a fixed size (number of points). The pickle is that I want to do that while keeping the intensity field.
First of all, in order to be able to access the intensity field, I load the pcd file through o3d.t.io.read_point_cloud(filename)

Downsampling:
I am currently doing the following:

pcdt = pcdt.farthest_point_down_sample(NUM_POINT)

However, it only returns a PointCloud object with x,y,z fields.

Upsampling
I am copying the fields pcdt.point['positions'] and pcdt.point['intensity'] into a numpy array of shape (N,4) and upsample it into an array of shape (N_wanted,4) and then reset the corresponding fields with the new values.
By doing this, Open3D warn me that it drops some fields in the process, however it keeps the fields that I want.
The problem comes when I want to write this new field into a .csv file while using the following function:

o3d.t.io.write_point_cloud(file_path, pcdt, write_ascii=True)

It write the file but drops the intensity. I don't know why since if I use it on the pcdt object before upsampling (and dropping the unwanted fields), it keeps all the fields in the new file.

I hope that I developed enough my situation. If not, just ask for complementary information :)

Thank you for your help!

PS: and I hope that I am not writing this in the wrong place (first time using this forum)

[saurabheights]

@destoswa Add a small reproducible code using available sample pcd from http://www.open3d.org/docs/latest/tutorial/data/index.html#pointcloud. It will allow me or others to quickly test.

[destoswa]

Okay so while writing a demo, I found the problem^^

So the problem that I had was that when the function o3d.t.io.write_point_cloud() write a PointCloude to a file, it will drop every attributes that haven't the right number of element (e.g. when the attributes intensity and positions are upsampled/downsampled, the attribute red will not have the same size anymore and will get dropped from the created .pcd file)
And during my process, I had a classical error which was that the intensity was not as a (N_new,1) shape but as a (N_new,) shape and was, therefore, dropped.

Regarding the down sampling, I simply used the function pcdt.farthest_point_down_sample(NUM_POINTS) to get the new set of points and took the corresponding intensity to recreate those two attribute.

Following is the working demo:

import open3d as o3d
import numpy as np
import random

def main():
    device = o3d.core.Device('CPU:0')
    dtype = o3d.core.float32
    num_duplicates = 50
    num_downsamp = 50

    # load pcd
    pcdt = o3d.t.io.read_point_cloud('sample.pcd')

    # remove duplicates
    point_set, idx = np.unique(pcdt.point['positions'].numpy(), axis=0, return_index=True)
    intensities = pcdt.point['intensity'].numpy()[idx].reshape(-1, 1)
    pcdt.point['positions'] = o3d.core.Tensor(point_set, dtype, device)
    pcdt.point['intensity'] = o3d.core.Tensor(intensities, dtype, device)

    # create copies
    pcdt_up = pcdt.clone()
    pcdt_down = pcdt.clone()
    
    # Upsample
    data = np.concatenate((pcdt.point['positions'].numpy(), pcdt.point['intensity'].numpy()), axis=1)
    lst_duplicates = random.sample(range(data.shape[0]), num_duplicates) # find a list of points that are going to be duplicated
    for el in lst_duplicates: # duplicate the points
        data = np.concatenate((data, data[el, :].reshape((1, -1))), axis=0)
    # update PointCloud and save file
    pcdt_up.point['positions'] = o3d.core.Tensor(data[:, :-1], dtype, device)
    pcdt_up.point['intensity'] = o3d.core.Tensor(data[:,-1].reshape(-1,1), dtype, device)
    o3d.t.io.write_point_cloud('sample_up.pcd', pcdt_up, write_ascii=True)
    
    # Downsample
    data = pcdt.farthest_point_down_sample(num_downsamp).point['positions'].numpy()
    intensities = pcdt.point['intensity'].numpy()
    old_pos = pcdt.point['positions'].numpy()
    mask = (old_pos == data[:, None]).all(2).any(0) # find which of the original pos has been kept
    intensities = intensities[mask] # apply the mask to keep only the intensities of the kept pos
    # update PointCloud and save file
    pcdt_down.point['positions'] = o3d.core.Tensor(data, dtype, device)
    pcdt_down.point['intensity'] = o3d.core.Tensor(intensities, dtype, device)
    o3d.t.io.write_point_cloud('sample_down.pcd', pcdt_down, write_ascii=True)


if __name__ == '__main__':
    main()

Thank you very much for making me create this demo, this is how I found the solution!

[saurabheights]

Welcome :). You can mark your answer as answer.