threecrate (Python)

Python bindings for threecrate — a high-performance 3D point cloud and mesh processing library written in Rust.

Installation

Pre-built wheels (no Rust required):

pip install threecrate

Build from source (requires Rust and maturin):

pip install maturin
cd threecrate-python
maturin develop --release

Quick Start

import numpy as np
import threecrate as tc

# Load a point cloud
cloud = tc.read_point_cloud("scan.ply")
print(cloud)  # PointCloud(120000 points)

# Or create from a numpy array (N, 3) float32
pts = np.random.rand(1000, 3).astype(np.float32)
cloud = tc.PointCloud.from_numpy(pts)

# Get points back as numpy
arr = cloud.to_numpy()  # shape (N, 3), dtype float32

API Reference

Types

Filtering

# Voxel grid downsampling
cloud = tc.voxel_downsample(cloud, voxel_size=0.05)

# Statistical outlier removal (default: k=20, std_ratio=2.0)
cloud = tc.remove_statistical_outliers(cloud, k_neighbors=20, std_ratio=2.0)

# Radius outlier removal
cloud = tc.remove_radius_outliers(cloud, radius=0.1, min_neighbors=5)

Normal Estimation

# Estimate normals using K nearest neighbours (default k=10)
normal_cloud = tc.estimate_normals(cloud, k_neighbors=10)

# Access positions and normals as numpy arrays
positions = normal_cloud.positions()  # (N, 3) float32
normals   = normal_cloud.normals()    # (N, 3) float32

Registration

# Point-to-point ICP
result = tc.icp(source, target, max_iterations=50)

print(result.converged)           # True / False
print(result.mse)                 # float
print(result.iterations)          # int
T = result.transformation()       # (4, 4) float32 numpy array

Segmentation

# RANSAC plane fitting
result = tc.segment_plane(cloud, threshold=0.01, max_iterations=1000)
coeffs = result.plane_coefficients()  # (4,) float32 [a, b, c, d]
indices = result.inlier_indices()     # list[int]
plane_cloud = result.inlier_cloud(cloud)   # PointCloud of inliers
print(result.num_inliers)

# Remove the dominant plane and keep the rest
non_plane_pts = [cloud.to_numpy()[i] for i in range(len(cloud))
                 if i not in set(indices)]

# Euclidean cluster extraction
clusters = tc.extract_clusters(cloud, tolerance=0.02,
                                min_cluster_size=100, max_cluster_size=25000)
for i, cluster in enumerate(clusters):
    print(f"Cluster {i}: {len(cluster)} points")

Mesh Simplification

# Reduce mesh to 50 % of original face count (quadric error decimation)
simplified = tc.simplify_mesh(mesh, reduction_ratio=0.5)
print(simplified.vertex_count, simplified.face_count)

Mesh Smoothing

# Laplacian smoothing (fast, mild shrinkage)
smooth = tc.smooth_mesh_laplacian(mesh, iterations=10, lambda_=0.5)

# Taubin smoothing (volume-preserving, recommended)
smooth = tc.smooth_mesh_taubin(mesh, iterations=10, lambda_=0.5, mu=-0.53)

# HC smoothing (good volume preservation with fine control)
smooth = tc.smooth_mesh_hc(mesh, iterations=10, alpha=0.0, beta=0.5)

Surface Reconstruction

# Automatic algorithm selection
mesh = tc.reconstruct(cloud)

# Poisson reconstruction (higher quality, requires normals)
normal_cloud = tc.estimate_normals(cloud)
mesh = tc.poisson_reconstruct(normal_cloud)

print(mesh.vertex_count)
print(mesh.face_count)
verts = mesh.vertices()  # (N, 3) float32
faces = mesh.faces()     # (M, 3) uint32

Mesh Health, Previews and 3D Printing

# Holes, pieces, flipped faces, watertight, size, area, volume, printable
report = tc.analyze_mesh(mesh)
print(report["printable"], report["problems"], report["size"], report["volume"])

# PNG from five views, no GPU needed. Orange = holes or inside-out faces.
tc.render([cloud, mesh], "check.png")
png = tc.render([mesh])                  # no path: returns the PNG bytes

# Point cloud or mesh in, watertight mesh out
printable, report = tc.repair_for_print(cloud, remove_floor=True, target_faces=200_000)
print(report["steps"], report["after"]["printable"])
tc.write_mesh(printable, "scan.printable.stl")

I/O

# Point clouds — PLY, PCD, XYZ, CSV, LAS, LAZ, E57
cloud = tc.read_point_cloud("scan.ply")
tc.write_point_cloud(cloud, "output.pcd")

# Meshes — PLY, OBJ, STL
mesh = tc.read_mesh("model.obj")
tc.write_mesh(mesh, "output.ply")

Full Example

import numpy as np
import threecrate as tc

# Load and preprocess
cloud = tc.read_point_cloud("scene.ply")
cloud = tc.voxel_downsample(cloud, voxel_size=0.02)
cloud = tc.remove_statistical_outliers(cloud, k_neighbors=20, std_ratio=2.0)

# Register two scans
source = tc.read_point_cloud("scan_a.ply")
target = tc.read_point_cloud("scan_b.ply")
result = tc.icp(source, target, max_iterations=100)
if result.converged:
    print(f"Aligned with MSE {result.mse:.4f}")
    print(result.transformation())

# Reconstruct surface
normal_cloud = tc.estimate_normals(cloud, k_neighbors=15)
mesh = tc.poisson_reconstruct(normal_cloud)
tc.write_mesh(mesh, "reconstruction.ply")
print(f"Mesh: {mesh.vertex_count} vertices, {mesh.face_count} faces")

Building from Source

Requirements: Rust 1.70+, Python 3.8+, maturin 1.x

# Install maturin
pip install maturin

# Development build (editable install)
cd threecrate-python
maturin develop --release

# Build a distributable wheel
maturin build --release --out dist/
pip install dist/threecrate-*.whl

License

Dual-licensed under MIT or Apache-2.0. See LICENSE-MIT for details.