[pcl::gpu::NormalEstimation] Unexpected Results

[yu-sheng-vincent]

Describe the bug

I was trying pcl::gpu::NormalEstimation() and found that the normal vectors act weird (point horizontally to the surface) under certain range of radius. I also ran pcl::NormalEstimationOMP() with the same searching radius and it went well.

Context

Generate normal surface using pcl::gpu::NormalEstimation().

Current Behavior

Part of the normal vectors point horizontally to the surface when running pcl::gpu::NormalEstimation(), while pcl::NormalEstimationOMP() works well with the same searching radius. Pictures are shown in the following section.

To Reproduce

To showcase the issue, I use this cutting blade object from Artec3D:
https://www.artec3d.com/3d-models/metal-cutting-blade

Screenshots/Code snippets

Snippet 1 - Using pcl::gpu::NormalEstimation()

#define SEARCH_RADIUS  (1)

int test_gpu(pcl::PointCloud<pcl::PointXYZ>::Ptr cloud, pcl::PointCloud<pcl::Normal>::Ptr cloud_normals)
{
  printf("Test with GPU\n");
  // Getinfo();

  typedef pcl::gpu::DeviceArray<pcl::PointXYZ>  g_PointXYZ;
  typedef pcl::gpu::DeviceArray<pcl::Normal>    g_Normal;

  g_PointXYZ dev_cloud(cloud->size());
  dev_cloud.upload(cloud->points.data(), cloud->size());
  
  printf("Normal Estimation...\n");
  pcl::gpu::NormalEstimation gne;
  gne.setInputCloud(dev_cloud);
  gne.setRadiusSearch(SEARCH_RADIUS * 100, 500);  // setRadiusSearch(float radius, int max_results);  What is max_results??
  gne.setViewPoint(-5, 5, 5);
  
  pcl::gpu::NormalEstimation::Normals dev_normals;
  gne.compute(dev_normals);

  std::vector<pcl::PointXYZ> normals;
  dev_normals.download(normals);
  
  //pcl::PointCloud<pcl::Normal>::Ptr cloud_normals (new pcl::PointCloud<pcl::Normal>);
  cloud_normals->resize(normals.size());
  for (size_t i = 0; i < normals.size(); ++i)
    {
      cloud_normals->points[i].normal_x = normals[i].x;
      cloud_normals->points[i].normal_y = normals[i].y;
      cloud_normals->points[i].normal_z = normals[i].z;
      cloud_normals->points[i].curvature = 0; // Curvature is not estimated in the GPU-based method
    }

  printf("Cloud Normal Size: %lu data points.\n", cloud_normals->size());

  return 0;
}

Snippet 2 - pcl::NormalEstimationOOP()

int test_pcl(pcl::PointCloud<pcl::PointXYZ>::Ptr cloud, pcl::PointCloud<pcl::Normal>::Ptr cloud_normals)
{
  printf("Test with CPU\n");

  // ==============================
  // Normal Estimation
  // ==============================
  pcl::NormalEstimationOMP<pcl::PointXYZ, pcl::Normal> ne;
  ne.setInputCloud(cloud);

  // KdTree to search K nearest
  pcl::search::KdTree<pcl::PointXYZ>::Ptr tree (new pcl::search::KdTree<pcl::PointXYZ> ());
  ne.setSearchMethod (tree);

  // parameters
  ne.setRadiusSearch(SEARCH_RADIUS);  // Consider all the points within the radius
  //ne.setKSearch(K_NEIGHBOR);  // mutex with setRadiusSearch
  ne.setViewPoint(-5, 5, 5);

  // place to store normal
  ne.compute(*cloud_normals);

  printf("Cloud Normal Size: %lu data points.\n", cloud_normals->size());

  return EXIT_SUCCESS;
}

Your Environment (please complete the following information):

• OS: Ubuntu 20.04
• Compiler: GCC 9.4.0
• PCL Version: 1.13.1
• CUDA Version: 12.1

[mvieth]

@yu-sheng-vincent Sorry for the late response. I tried to reproduce the problem, but the results I get seem correct. Could you share your whole code? Especially how you save the cloud normals to a file? You use meshlab to visualize the cloud and normals, is that correct?

[mvieth]

We found the problem, please see here: #5846