Efficient point cloud cropper

Efficient 3D point cloud bounding-box cropping. Implemented with parallel stream compaction using TBB's parallel_scan.

Installation

# Install CMake
sudo apt install cmake # Install globally
conda install cmake    # Install in our conda env

# Get code.
git clone https://github.com/yxlao/pointcloud_cropper
cd pointcloud_cropper

# Activate your virtualenv.
# Build and install.
python setup.py install

# Test.
python benchmark.py

Usage

import pointcloud_cropper as pc
import numpy as np

# Prepare (N, 3) points of np.float64 or np.float32.
np.random.seed(0)
points = np.random.rand(int(1e8), 3)
print(points.shape)

# get_crop_indices returns the selected index.
indices_np = pc.get_crop_indices(points=points,
                                 x_min=0.25,
                                 y_min=0.25,
                                 z_min=0.25,
                                 x_max=0.75,
                                 y_max=0.75,
                                 z_max=0.75)
cropped_points = points[indices_np]
print(cropped_points.shape)

Performance

1e8 points, 50% selection ratio. Measured with Intel i9-9980XE.

Time with numpy: 1.3744 sec
Time with TBB  : 0.1031 sec