Skip to content

Features Extraction

features_extraction: Extract geometric features from point clouds: normals, curvature, and FPFH descriptors for registration and analysis.
Methods normal_estimation, fpfh_features, curvature_estimation
Input files .ply, .pcd, .txt, .xyz
Bundled datasets to try open3d_bunny, open3d_icp_demo, kitti_raw_2011_09_26_drive_0001, medshapenet, c3vd_endoscopy_splats
CLI pardesline run features_extraction.<method> input.ext [--param value]

How it works

Geometric feature extraction computes descriptors that characterize the local or global shape of a point cloud.

Normal estimation fits a local tangent plane to neighbors. Curvature is estimated from eigenvalues of the local covariance matrix. FPFH descriptors encode angular relationships between normals.

Methods

normal_estimation

The normal \(\vec{n}_i\) is the eigenvector corresponding to the smallest eigenvalue of \(\mathbf{C}_i\). Normals are oriented consistently using tangent plane propagation.

import pardesline as pl

obj = pl.load("input.ply")
result = obj.run("features_extraction.normal_estimation")  # catalog defaults
# same thing, domain vocabulary: `points.normals()`
result.provenance[-1]  # the step, every parameter used, duration, version

Output: .ply. Script: python -m pardesline.ops.features_extraction.normal_estimation --input ... --output ....

Defaults validated on: Stanford bunny through the registration pipeline; generated clouds in tests/test_ops_smoke.py

Parameter Type Default Range What it does Why this default, when to change it
radius float 0.05 0.005 to 0.5 Search radius for normal estimation. 0.05 in file units: about 5x the spacing of a centimetre-resolution scan; the covariance needs 10-30 points inside the radius.
max_nn int 30 10 to 100 Maximum nearest neighbors. 30 neighbours cap the cost and keep normals local at edges.
fast_normal_computation (advanced) bool true Use the faster (less accurate) covariance eigen solver in estimate_normals. On: Open3D's closed-form eigen solver is accurate to visual precision and several times faster than the iterative one.
orient_k (advanced) int 0 0 to 100 k neighbors for consistent tangent-plane orientation (0 = auto min(max_nn,15)). 0 lets the script choose min(max_nn, 15) for the orientation propagation: small enough to stay consistent across thin parts.

fpfh_features

Fast Point Feature Histograms encode angular relationships between normals into a 33-dimensional descriptor using three angular features \((\alpha, \phi, \theta)\).

import pardesline as pl

obj = pl.load("input.ply")
result = obj.run("features_extraction.fpfh_features")  # catalog defaults
result.provenance[-1]  # the step, every parameter used, duration, version

Output: .ply. Script: python -m pardesline.ops.features_extraction.fpfh_features --input ... --output ....

Defaults validated on: Stanford bunny and Open3D fragments through registration.ransac_global; generated clouds in tests/test_ops_smoke.py

Parameter Type Default Range What it does Why this default, when to change it
voxel_size float 0.005 0.001 to 0.05 Voxel size for downsampling before feature computation. 0.005 in file units (5 mm on a metre object) keeps enough points for descriptive histograms; the registration pipeline derives it from the bounding box instead.
normal_radius_factor (advanced) float 2.0 1.0 to 10.0 Normal search radius as a multiple of voxel_size. 2x the voxel for normals.
normal_max_nn (advanced) int 30 10 to 200 Maximum nearest neighbors for normal estimation. 30 neighbours cap the normal-estimation cost on dense clouds while keeping the covariance stable.
feature_radius_factor (advanced) float 5.0 1.0 to 20.0 FPFH search radius as a multiple of voxel_size. 5x the voxel: the Open3D-recommended FPFH support.
feature_max_nn (advanced) int 100 10 to 500 Maximum nearest neighbors for FPFH feature computation. 100 neighbours bound the histogram cost.

curvature_estimation

Curvature \(\kappa_i = \frac{\lambda_0}{\lambda_0 + \lambda_1 + \lambda_2}\). Near 0 = flat, near \(1/3\) = high curvature.

import pardesline as pl

obj = pl.load("input.ply")
result = obj.run("features_extraction.curvature_estimation")  # catalog defaults
result.provenance[-1]  # the step, every parameter used, duration, version

Output: .ply. Script: python -m pardesline.ops.features_extraction.curvature_estimation --input ... --output ....

Defaults validated on: generated clouds in tests/test_ops_smoke.py

Parameter Type Default Range What it does Why this default, when to change it
radius float 0.05 0.005 to 0.5 Search radius for curvature computation. 0.05 in file units for the covariance neighbourhood, as for normals.
max_nn int 30 10 to 100 Maximum nearest neighbors. 30 neighbours cap the cost and keep the estimate local at edges.
normalize_percentile (advanced) float 99 50 to 100 Percentile of curvature used as the color-normalization ceiling. Colour scale capped at the 99th percentile so a few extreme points (edges, noise) do not flatten the map.

Generated from src/pardesline/catalog/data/features_extraction.json and _defaults.json on 2026-10-10.