Skip to content

Point Cloud Segmentation

point_cloud_segmentation: Segment a point cloud into meaningful regions using geometric methods: RANSAC plane fitting and DBSCAN clustering.
Methods ransac_plane, dbscan_clustering
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 point_cloud_segmentation.<method> input.ext [--param value]

How it works

Point cloud segmentation partitions a set of 3D points \(P = \{p_1, \ldots, p_n\}\) into disjoint subsets corresponding to semantic or geometric regions.

RANSAC-based plane fitting iteratively fits a plane model \(\vec{n} \cdot \vec{x} + d = 0\) to random point subsets.

DBSCAN groups points based on density: a point is a core point if it has at least \(n_{\min}\) neighbors within radius \(\epsilon\).

Methods

ransac_plane

RANSAC iteratively fits a plane model and identifies inliers within distance threshold \(\epsilon\). Plane inliers are colored distinctly from outliers.

import pardesline as pl

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

Output: .ply. Script: python -m pardesline.ops.point_cloud_segmentation.ransac_plane --input ... --output ....

Defaults validated on: generated clouds in tests/test_ops_smoke.py; the same RANSAC plane model runs on KITTI in lidar_scene_detection

Parameter Type Default Range What it does Why this default, when to change it
distance_threshold float 0.01 0.001 to 0.1 Maximum distance from point to plane for inlier. 0.01 in file units: one centimetre on metre-scale indoor scans, which is the depth-camera noise level; use 0.2-0.3 m for outdoor LiDAR ground.
num_iterations int 1000 100 to 10000 Number of RANSAC iterations. 1 000 iterations find the dominant plane with probability above 0.99 when it holds at least 20 % of the points.
ransac_n (advanced) int 3 3 to 10 Number of points sampled to estimate each plane candidate. 3 points define a plane; the minimum keeps the inlier probability per draw highest.
probability (advanced) float 0.99999999 0 to 1 Expected probability of finding the best plane (early-stop confidence). 0.99999999 lets Open3D stop early once the best plane is almost certainly found, so the 1 000 iterations are a ceiling.

dbscan_clustering

DBSCAN groups points based on density. Given parameters \((\epsilon, n_{\min})\), connected core points form clusters. Each cluster gets a unique color.

import pardesline as pl

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

Output: .ply. Script: python -m pardesline.ops.point_cloud_segmentation.dbscan_clustering --input ... --output ....

Defaults validated on: generated clouds in tests/test_ops_smoke.py; KITTI sweeps through lidar_scene_detection.object_clustering

Parameter Type Default Range What it does Why this default, when to change it
eps float 0.02 0.001 to 0.5 Maximum distance between two points in same cluster. 0.02 in file units: two centimetres joins points of the same object on indoor depth scans; use 0.5 m outdoors (see object_clustering).
min_points int 10 3 to 100 Minimum points to form a cluster. 10 points keep speckle out of the cluster list while small objects (a cup, a hand) still form one.
print_progress (advanced) bool false Print an Open3D progress bar during clustering. Off: terminal output only.

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