Filtering pointcloud

Question

I wish to filter a pointcloud, loaded with opend3d , as efficiently as possible.

Currently, I perform a downsampling of the points before making a mesh out of them and using .contains on an inclusion volume mesh I did manually. Something like this:

    def load_pointcloud(self, pointcloud_path): 
        # Load Pointcloud
        print('target_pointcloud', pointcloud_path)
        self.pointcloud_path = pointcloud_path

        pcd = o3d.io.read_point_cloud(pointcloud_path)
        downpcd = pcd.voxel_down_sample(voxel_size=0.02)
        cl, ind = downpcd.remove_statistical_outlier(nb_neighbors=20,
                                                std_ratio=2.0)
        downpcd = downpcd.select_by_index(ind)
        pcd_points = np.asarray(downpcd.points, dtype=np.float32)

        self.verts = torch.from_numpy(pcd_points)
        self.verts = self.verts.to(device)

        # We construct a Meshes structure for the target mesh
        self.pointcloud_points = Pointclouds(points=[self.verts])
        self.points = pcd_points
        self.inclusion_pointcloud()

    def inclusion_pointcloud(self):
        vetices_in_mesh_states = self.mesh_inclusion.contains(self.points)
        vetices_in_mesh = self.points[vetices_in_mesh_states == True]

        # Creating cropped point cloud
        cropped_pc = o3d.geometry.PointCloud()
        cropped_pc.points = o3d.utility.Vector3dVector(vetices_in_mesh)
        cropped_pc.paint_uniform_color([0,0,0])

        self.points = np.asarray(cropped_pc.points, dtype=np.float32)
        self.verts = torch.from_numpy(self.points)
        self.verts = self.verts.to(device)
        self.pointcloud_points = Pointclouds(points=[self.verts])
        self.pc_mesh = trimesh.Trimesh(vertices=self.points)  

What I was thinking in doing was to, after the downsampling, mask away points on X, Y, and Z, and then making a mesh to use .contains again in the same inclusion volume. I thought that this would reduce the .contains computation and run faster, and it kind of does, but is a marginal reduction, like 10 or 15ms, sometimes less. Something like this:

    def new_load_pointcloud(self, pointcloud_path): 
        # Load Pointcloud
        print('target_pointcloud', pointcloud_path)
        self.pointcloud_path = pointcloud_path

        pcd = self.trim_cloud(pointcloud_path)
        downpcd = pcd.voxel_down_sample(voxel_size=0.02)
        cl, ind = downpcd.remove_statistical_outlier(nb_neighbors=20,
                                                std_ratio=2.0)
        downpcd = downpcd.select_by_index(ind)
        pcd_points = np.asarray(downpcd.points, dtype=np.float32)

        self.verts = torch.from_numpy(pcd_points)
        self.verts = self.verts.to(device)

        # We construct a Meshes structure for the target mesh
        self.pointcloud_points = Pointclouds(points=[self.verts])
        self.points = pcd_points
        self.inclusion_pointcloud()

    def trim_cloud(self, pointcloud_path):
        # pcd = o3d.io.read_point_cloud(pointcloud_path)
        pcd_clean = o3d.io.read_point_cloud(pointcloud_path)

        # X Axis
        points = np.asarray(pcd_clean.points)
        mask_x_1 = points[:,0] > -0.4
        mask_x_2 = points[:,0] < 0.4

        # Y Axis
        mask_y_1 = points[:,1] > -1.3
        mask_y_2 = points[:,1] < 0.9

        # Z Axis
        mask_z_1 = points[:,2] < 0.3 # Closer to floor     
        mask_z_2 = points[:,2] > -0.1 # Clooser to ceiling

        mask_x = np.logical_and(mask_x_1, mask_x_2) # Along table's wide
        mask_y = np.logical_and(mask_y_1, mask_y_2) # Along table's longitude
        mask_z = np.logical_and(mask_z_1, mask_z_2) # Along table's height
        mask = np.logical_and(mask_x, mask_y, mask_z)
        pcd_clean.points = o3d.utility.Vector3dVector(points[mask])

        return pcd_clean

    def inclusion_pointcloud(self):
        vetices_in_mesh_states = self.mesh_inclusion.contains(self.points)
        vetices_in_mesh = self.points[vetices_in_mesh_states == True]

        # Creating cropped point cloud
        cropped_pc = o3d.geometry.PointCloud()
        cropped_pc.points = o3d.utility.Vector3dVector(vetices_in_mesh)
        cropped_pc.paint_uniform_color([0,0,0])

        self.points = np.asarray(cropped_pc.points, dtype=np.float32)
        self.verts = torch.from_numpy(self.points)
        self.verts = self.verts.to(device)
        self.pointcloud_points = Pointclouds(points=[self.verts])
        self.pc_mesh = trimesh.Trimesh(vertices=self.points)  

Answer 1

I think you are using too much nb_neighbors for the filter. Try less points, like 6 or 10, and a better threshold, like 1.0 or even 0.5. Here is the same filter on MATLAB documentation https://www.mathworks.com/help/vision/ref/pcdenoise.html , standard value for the threshold is 1.0 and for the knn is 6. You can also try the Radius Outlier Removal or the median filter: https://www.mathworks.com/help/lidar/ref/pcmedian.html