Point Cloud Feature Analysis

Preprocessing and FPFH feature analysis pipeline on the Stanford Bunny dataset, built with Python and Open3D.

What it does

Runs a 6-stage preprocessing pipeline on the Stanford Bunny .ply file, then performs two types of geometric feature analysis using FPFH (Fast Point Feature Histograms).

Preprocessing pipeline:

Stage	Output
Load	35,947 points from raw scan
Voxel downsampling	3,023 points
Statistical outlier removal	2,879 points
Normal estimation	Normals computed via KDTree
FPFH extraction	(33, 2879) feature matrix
Visualization	Point cloud + normals

Feature analysis:

• Per-point feature coloring — maps a single FPFH bin across all points to a plasma colormap, revealing which surface regions share geometric character
• Geometric clustering — K-means (k=5) on the full 33-dimensional FPFH vectors groups points by local geometry, independent of spatial position

What is FPFH?

FPFH (Fast Point Feature Histograms) describes the local geometry around each point as a 33-bin histogram. It captures how surface normals vary across a point's neighborhood — flat regions, edges, and corners each produce a distinct histogram shape. Points with similar FPFH vectors have similar local geometry regardless of where they are on the surface.

Results

Normals

Feature coloring — surface colored by FPFH bin 22 values

Geometric clustering — 5 clusters by FPFH similarity

Setup

git clone <repo>
cd pointcloud_feature_analysis
python3 -m venv venv
source venv/bin/activate
pip install -r requirements.txt

Download the Stanford Bunny from graphics.stanford.edu/data/3Dscanrep and place bun_zipper.ply at:

data/bunny/reconstruction/bun_zipper.ply

Run

XDG_SESSION_TYPE=x11 python3 src/pipeline.py

XDG_SESSION_TYPE=x11 is required on Ubuntu 24.04 with Wayland — Open3D's visualizer needs an X11 context.

Stack

• Python 3
• Open3D 0.19.0
• NumPy 2.4.4
• scikit-learn