about farthest point sampling

[amiltonwong]

Hi, authors,

The Farthest Point Sampling (FPS) is applied in the official implementation of pointnet2. However, I could not find FPS implementation in your package? Does your package provide FPS function?

THX!

[erikwijmans]

https://github.com/erikwijmans/Pointnet2_PyTorch/blob/master/pointnet2/_ext-src/src/sampling_gpu.cu#L70

[Leerw]

Hi,
@erikwijmans
I met a wrong result after applying FPS,
the data I used is ShapeNetCore.v1/02691156/52a1b6e8177805cf53a728ba6e36dfae/model.obj,
after sampling this model and apply fps, there is a hole in the point cloud.
Please check it.

[erikwijmans]

Can you please provide screenshots or something? Also note that FPS algorithm used in PointNet++ is a) approximate (point zero is always included in the output set) and b) greedy, meaning that it is completely liable to make some errors.

[Leerw]

test_data.zip

You can get the test data from this .zip file, including a mesh model.obj, a point cloud model.pcd.
I sampled 16384 points on this mesh to get a point cloud, afterwards, I apply your fps method, and get a point cloud with a hole in the middle of the airplane, as showed in this picture.

test code:

import numpy as np
import open3d as o3d
import torch
from pointnet2_ops.pointnet2_utils import furthest_point_sample, gather_operation

resample = lambda points, n: gather_operation(points.transpose(1 , 2).contiguous(), furthest_point_sample(points, n))


pcd = o3d.io.read_point_cloud("./ShapeNetCore.v1/02691156/52a1b6e8177805cf53a728ba6e36dfae/model.pcd")

points = torch.from_numpy(np.asarray(pcd.points).astype(np.float32)).cuda().unsqueeze(0)
print(points.shape, points.dtype)
points = resample(points, 5000)

def vis(points):
    points = points.transpose(1, 2).contiguous().cpu().numpy()[0]
    pcd = o3d.geometry.PointCloud()
    pcd.points = o3d.utility.Vector3dVector(points)
    o3d.visualization.draw_geometries([pcd])

vis(points)

[erikwijmans]

Can you try removing these two lines: https://github.com/erikwijmans/Pointnet2_PyTorch/blob/master/pointnet2_ops_lib/pointnet2_ops/_ext-src/src/sampling_gpu.cu#L100-L101

I filtered out near-zero points to avoid some numerical instabilities, but, if a model is very small, that may cause issue.

[Leerw]

Can you try removing these two lines: https://github.com/erikwijmans/Pointnet2_PyTorch/blob/master/pointnet2_ops_lib/pointnet2_ops/_ext-src/src/sampling_gpu.cu#L100-L101

I filtered out near-zero points to avoid some numerical instabilities, but, if a model is very small, that may cause issue.

Thanks, this issue fixed.

[erikwijmans]

Sounds good. I will consider removing that or making the limit tighter. IIRC, the instability was due to my specific down-stream application at the time.

[pixar0407]

Can you please provide screenshots or something? Also note that FPS algorithm used in PointNet++ is a) approximate (point zero is always included in the output set) and b) greedy, meaning that it is completely liable to make some errors.

hi, @erikwijmans

Thank you for the quality code. I have few questions.

1. you mentioned 'FPS algorithm used in PointNet++'. So you mean that the algorithm who have implemented here(https://github.com/erikwijmans/Pointnet2_PyTorch/blob/master/pointnet2_ops_lib/pointnet2_ops/_ext-src/src/sampling_gpu.cu) also has those two a), b) properties?

2. And can you elaborate the meaning of 'greedy'?

3. Efficiency of your FPS algorithm
   i have tried two different version of FPS algorithm, yours and this one pytorch_cluster.

And yours seems pretty fast. (same condition, 1024 points to 512 points, etc)
Can you tell me any reason why your code performs better?