Parametric plane detection from 3D point cloud via probabilistic clustering

Abstract

A plane is the most fundamental isotropic geometric shape commonly found in human-made environments. Plane detection serves as a bridge between low-level 3D raw data and high-level structural representations of underlying 3D shapes. Existing methods often require extensive parameter tuning, rely on normal information, or depend on large training datasets, limiting their generalization across diverse shapes. In this paper, we introduce a novel algorithm for extracting high-quality planar information from 3D point clouds, inspired by unsupervised probabilistic clustering. Our approach reformulates plane detection as a Bayesian probabilistic modeling problem within a machine learning framework. By theoretically deriving a maximum likelihood function for plane parameters, we leverage the classic yet powerful expectation-maximization (EM) algorithm to iteratively solve for maximum likelihood estimates. Our framework is highly flexible, enabling the accurate extraction of parametric planes from various scenes and sensor types. Extensive experiments and comparisons with both analytical and learning-based methods demonstrate the superiority of our approach. Moreover, our algorithm is simple, efficient, and easy to implement. The code is released to facilitate future research: https://github.com/dengdaigithub/PPD.