Pe3d

Pe3d is a Python 3.11 command-line tool for extracting pebble or grain metrics from 3D point clouds.

It is a Python fork and reimplementation of the G3Point workflow, packaged as a standalone CLI.

It provides an end-to-end workflow for:

• preprocessing .ply, .las, and .laz point clouds,
• watershed-style grain segmentation,
• grain merging and cleaning,
• cuboid and ellipsoid fitting,
• grain-size and orientation statistics,
• export of per-grain .ply files and summary tables,
• interactive manual block / non-block labeling for evaluation datasets.

Installation

pip install pe3d

Quick Start

pe3d run --pointcloud path/to/cloud.ply

Optional arguments:

• --param-csv: dataset parameter table matching the MATLAB-style param.csv format
• --output-root: output directory root, default pe3d_outputs
• --sizing-mode: ellipsoid for rounded grains or block for irregular protruding blocks
• --fit-strategy: pca_maxabs, bounded_pca, or hybrid_direct
• --orientation-mode: atan2 or matlab
• --preprocess-mode: matlab or fast
• --ground-mode: none or smrf for an optional morphological ground/background filter in block mode
• --smrf-cell-size: SMRF grid cell size, use 0 for automatic sizing from point spacing
• --smrf-max-window: maximum SMRF opening window in cloud units
• --smrf-height-threshold: base height threshold above the SMRF surface
• --smrf-slope-threshold: slope-dependent additive threshold multiplier for SMRF
• --max-ellipsoid-cloud-fraction: reject ellipsoids larger than this fraction of the processed cloud max span, default 0.25, use 0 to disable
• --workers: per-grain fitting worker threads, default 1, use 0 for auto
• --save-grains or --no-save-grains: export per-grain .ply files, default off
• --save-colored-clouds or --no-save-colored-clouds: export the original point cloud recolored by ellipsoid volume, elongation, and azimuth, default off
• --block-split or --no-block-split: experimental local watershed splitter for oversized merged block components, default off
• --skip-plots

Example:

pe3d run \
  --pointcloud data/sample_cloud.laz \
  --param-csv data/param.csv \
  --sizing-mode block \
  --preprocess-mode matlab \
  --ground-mode smrf \
  --max-ellipsoid-cloud-fraction 0.25 \
  --no-block-split \
  --workers 1 \
  --save-colored-clouds \
  --save-grains \
  --output-root results

ellipsoid mode is the original G3Point-style workflow and is best suited to rounded pebbles or grains.

block mode is an alternative backend for irregular debris or avalanche deposits. It combines local relief, local vertical spread, local scattering, and multiscale PCA curvature / planarity into a blockiness score, segments connected components from that score, then measures each object with oriented bounding-box axes and convex-hull volume/area instead of relying on an ellipsoid fit. --ground-mode smrf adds a simple morphological ground/background filter before segmentation. The optional --block-split stage adds a local map-view watershed inside only the largest merged components. It remains disabled by default because it was tuned on Nuage_HP manual instances and has not yet been validated on other datasets.

Manual Labeling

Use pe3d label to open an interactive labeling window for manual block annotation:

pe3d label \
  --pointcloud data/sample_cloud.ply \
  --output-root manual_labels \
  --reset-session \
  --max-display-points 120000 \
  --rotate-horizontal

The tool fits a dominant plane, rotates the cloud into a horizontal map view, and opens:

• a 3D inspection pane,
• a top-down map view where polygon selection happens.

The simplified session model stores one polygon for block and one polygon for non-block. Drawing a new polygon for a mode replaces the previous one. Clear Mode removes the currently selected zone.

Use --reset-session to ignore and delete any previously saved polygons for that dataset.

The session saves:

• *_manual_labels.npz: point-wise label array,
• *_manual_labels.json: metadata and counts,
• *_manual_polygons.json: stored block / non-block polygons,
• *_manual_labels.ply: RGB preview cloud for CloudCompare or similar viewers.

Use pe3d label-instances when you want to isolate individual blocks instead of broad zones:

pe3d label-instances \
  --pointcloud data/sample_cloud.ply \
  --output-root manual_labels \
  --reset-session \
  --max-display-points 120000 \
  --rotate-horizontal

In this mode, each polygon creates one block instance id. The session saves:

• *_manual_instances.npz: point-wise integer instance labels,
• *_manual_instances.json: metadata and counts,
• *_manual_instances_polygons.json: one polygon per saved instance,
• *_manual_instances.ply: RGB preview cloud with one color per instance.

Performance Note

Pe3d uses multithreaded cKDTree queries internally, and per-grain fitting can also be threaded with --workers.

In the current benchmark on this repository, serial fitting (--workers 1) was faster than auto-threaded fitting (--workers 0) on both Otira and Nuage_HP, while producing identical CSV outputs. See:

• evaluation/worker_benchmark/README.md

Validation Snapshot

This repository includes a small public evaluation snapshot comparing the current Python implementation to archived reference outputs derived from the original workflow.

On the retained Otira comparison snapshot, the current Python release candidate matches 396 reference rows with:

• Dmax MAE: 0.1289 m
• Dmed MAE: 0.0894 m
• Dmin MAE: 0.0781 m
• angle_Mview MAE: 37.62 deg
• angle_Xview MAE: 31.82 deg

That snapshot indicates stronger size parity than orientation parity, which is the main remaining gap relative to the MATLAB workflow.

Inputs and Outputs

Input point clouds can be .ply, .las, or .laz and must provide x, y, and z coordinates. Optional RGB fields are preserved on load when present. Per-grain exports remain .ply.

Each run creates:

• Figure/<dataset>_nX/
• Grain/<dataset>_nX/
• Excel/<dataset>_nX/
• Cloud/<dataset>_nX/

The summary table is exported as both CSV and XLSX.

When --save-colored-clouds is enabled, Cloud/<dataset>_nX/ contains three RGB .ply files built from the original point set:

• *_by_volume.ply
• *_by_elongation.ply
• *_by_azimuth.ply

When --ground-mode smrf is active, the same folder also contains:

• *_height_above_ground.ply
• *_ground_mask.ply

License and Attribution

Pe3d is currently distributed under the same license terms as the original G3Point MATLAB code from which this Python fork is derived.

That means redistribution and modification are permitted, but the original copyright notice, conditions, and disclaimer must be retained. The original license also requires scientific or technical publications that benefit from the software to acknowledge its use and cite the corresponding publication.

See:

• LICENSE
• NOTICE.md

Attribution and citation source:

• G3Point software page: https://lidar.univ-rennes.fr/en/g3point
• Steer, Guérit, Lague, Crave, and Gourdon (2022), https://doi.org/10.5194/esurf-10-1211-2022