如何使用trimesh从无法线/顶点顺序的点云创建曲面网格？

Below are practical alternatives to convex hull for generating triangle meshes from point clouds using trimesh, including fixes for planar point cloud cases where convex hull fails:

1. Planar Point Cloud Triangulation with Scipy Delaunay #

Convex hull only captures the outer boundary of planar points—use Scipy's Delaunay triangulation to generate a full mesh of the planar point set:

import trimesh
import numpy as np
from scipy.spatial import Delaunay

# Replace with your planar point cloud data
points = np.random.rand(100, 3)
points[:, 2] = 0.0  # Example: planar along z-axis

# Identify and drop the constant axis to get 2D points
axis = np.argmin(np.var(points, axis=0))
points_2d = np.delete(points, axis, axis=1)

# Run Delaunay triangulation
tri = Delaunay(points_2d)

# Create trimesh mesh and compute normals
mesh = trimesh.Trimesh(vertices=points, faces=tri.simplices)
mesh.compute_normals()

# Validate and visualize
print(f"Generated {len(mesh.faces)} faces")
mesh.show()

2. Ball Pivoting via Open3D + Trimesh Conversion #

Since trimesh lacks a native Ball Pivoting implementation, use Open3D's proven method and convert the result to trimesh format:

import trimesh
import open3d as o3d
import numpy as np

# Replace with your point cloud data
points = np.random.rand(200, 3)

# Convert to Open3D PointCloud and estimate normals
pcd = o3d.geometry.PointCloud()
pcd.points = o3d.utility.Vector3dVector(points)
pcd.estimate_normals(search_param=o3d.geometry.KDTreeSearchParamHybrid(radius=0.1, max_nn=30))

# Run Ball Pivoting (adjust radii based on point density)
radii = [0.05, 0.1, 0.2]
mesh_o3d = o3d.geometry.TriangleMesh.create_from_point_cloud_ball_pivoting(pcd, o3d.utility.DoubleVector(radii))

# Convert Open3D mesh to trimesh
mesh_trimesh = trimesh.Trimesh(
    vertices=np.asarray(mesh_o3d.vertices),
    faces=np.asarray(mesh_o3d.triangles),
    vertex_normals=np.asarray(mesh_o3d.vertex_normals)
)

mesh_trimesh.show()

3. Screened Poisson Reconstruction via Open3D + Trimesh Conversion #

For dense point clouds needing smooth surfaces, use Open3D's Poisson reconstruction and convert to trimesh:

import trimesh
import open3d as o3d
import numpy as np

# Replace with your dense point cloud data
points = np.random.rand(500, 3)

# Convert to Open3D PointCloud and estimate normals
pcd = o3d.geometry.PointCloud()
pcd.points = o3d.utility.Vector3dVector(points)
pcd.estimate_normals(search_param=o3d.geometry.KDTreeSearchParamHybrid(radius=0.1, max_nn=30))

# Run Poisson reconstruction (adjust depth for detail)
mesh_o3d, densities = o3d.geometry.TriangleMesh.create_from_point_cloud_poisson(pcd, depth=8)

# Optional: Clean up low-density vertices
vertices_to_remove = densities < np.quantile(densities, 0.1)
mesh_o3d.remove_vertices_by_mask(vertices_to_remove)

# Convert to trimesh
mesh_trimesh = trimesh.Trimesh(
    vertices=np.asarray(mesh_o3d.vertices),
    faces=np.asarray(mesh_o3d.triangles),
    face_normals=np.asarray(mesh_o3d.triangle_normals)
)

mesh_trimesh.show()

Key Tips #

• For planar data: Stick to Delaunay triangulation for efficiency and accuracy.
• Ball Pivoting: Best for sparse point clouds with distinct surface edges; tune radii to match your point spacing.
• Poisson Reconstruction: Ideal for dense data to produce smooth surfaces; adjust depth to control mesh detail.
• Always validate normals with mesh.compute_normals() if they're missing or incorrect.

内容的提问来源于stack exchange，提问作者maurock