🌐 SHOT / 📊 FPFH descriptors on 3D Point Clouds

This repository contains resources to compute SHOT and FPFH, two 3D descriptors on point clouds. The main script main.py contains a pipeline consisting of the following steps:

• Data retrieving: binary .ply files are supported. Two point clouds will be loaded.
• Query point selection: an algorithm taken from query_points_selection.py will be used to select a subset of query points.
• Descriptors computation: descriptors will be computed on the query points.
• Descriptors matching: descriptors between the two point clouds will be matched using one of the algorithms found in matching.py and a coarse registration will be performed.
• Fine registration: an ICP-based registration will be performed.
• Data saving: if not disabled, 4 point clouds will be saved: 2 for each descriptor, one for the coarse registration and another one for the fine registration.

💬 Basic usage

Run python main.py --help to display a help message specifying the arguments of this main script. This script will run the steps specified above in the same order, each argument finding a role in one of the steps.

📦 Dependency management

Dependencies are managed using poetry: https://python-poetry.org/

🔧 Implementation

The descriptors are implemented from scratch based on NumPy. The KDTree implementation used here is the one from scikit-learn.neighbors.

Possible performance-improving extensions

• Using numba to jit the code. The compilation is quite long and the jitting fails on various functions like KDTree.query_radius or np.mean (https://stackoverflow.com/questions/57500001/numba-failure-with-np-mean). It could work with some additional work to refactor the code in a numba-friendly manner.

• Use the KDTree from scipy instead of scikit-learn to allow for parallel computing on several workers at once.