[Paper Review] Semantic Classification of 3D Point Clouds with Multiscale Spherical Neighborhoods

Hugues Thomas, Jean‐Emmanuel Deschaud+3 more|arXiv (Cornell University)|Aug 1, 2018

3D Shape Modeling and Analysis23 references4 citations

TL;DR

This paper proposes a novel multiscale spherical neighborhood definition for 3D point cloud semantic classification, using proportional subsampling to ensure geometric consistency and feature stability. The method achieves state-of-the-art performance with random forests, outperforming both traditional handcrafted features and deep learning models on large-scale datasets, demonstrating that well-designed geometric features can rival complex learning-based approaches without segmentation or post-processing.

ABSTRACT

This paper introduces a new definition of multiscale neighborhoods in 3D point clouds. This definition, based on spherical neighborhoods and proportional subsampling, allows the computation of features with a consistent geometrical meaning, which is not the case when using k-nearest neighbors. With an appropriate learning strategy, the proposed features can be used in a random forest to classify 3D points. In this semantic classification task, we show that our multiscale features outperform state-of-the-art features using the same experimental conditions. Furthermore, their classification power competes with more elaborate classification approaches including Deep Learning methods.

Motivation & Objective

• To address the lack of geometric consistency in k-nearest neighbor (KNN) based multiscale neighborhoods, which distort features across scales.
• To develop a multiscale neighborhood definition that preserves geometric meaning while enabling consistent feature computation across varying point densities.
• To evaluate whether handcrafted geometric features, derived from the new neighborhood definition, can match or exceed the performance of deep learning-based semantic segmentation methods.
• To demonstrate generalization across diverse datasets and scanning technologies, including airborne and terrestrial lidar.
• To establish a robust, efficient, and interpretable alternative to deep learning for 3D semantic classification.

Proposed method

• Proposes a multiscale spherical neighborhood definition based on fixed-radius spherical neighborhoods, where each scale is derived via proportional subsampling to maintain consistent point density.
• Uses iterative subsampling to generate multiple scales, ensuring that each neighborhood contains a controlled number of points (parameterized by ρ) while preserving spatial structure.
• Computes simple geometric features—such as local covariance, verticality, and height distribution—within each spherical neighborhood to describe local geometry.
• Applies a random forest classifier to predict semantic labels for each point independently, using the multiscale features as input.
• Employs a cross-city training and testing protocol (e.g., train on Lille, test on Paris) to evaluate generalization and robustness.
• Uses a parameter ρ to control the number of subsampled points per neighborhood, balancing feature quality and computational efficiency.

Experimental results

Research questions

• RQ1 Can a multiscale spherical neighborhood definition with proportional subsampling improve the geometric consistency and discriminative power of 3D point cloud features compared to KNN-based neighborhoods?
• RQ2 To what extent can handcrafted geometric features, derived from the proposed neighborhood definition, outperform state-of-the-art features in the same experimental setup?
• RQ3 Can a simple random forest classifier using these features achieve performance competitive with complex deep learning-based semantic segmentation models?
• RQ4 How does the method generalize across different cities, scanning technologies, and point cloud densities?
• RQ5 What is the trade-off between feature quality and computational speed when varying the subsampling parameter ρ?

Key findings

• The proposed multiscale spherical neighborhoods significantly improve feature consistency and geometric meaning compared to KNN-based neighborhoods, which suffer from distortion at varying scales.
• On the Semantic3D and Paris-Lille-3D datasets, the method outperforms state-of-the-art handcrafted features under identical experimental conditions, achieving higher mean Intersection over Union (mIoU).
• On the larger Paris-Lille-3D dataset, the method achieves competitive mIoU scores that rival or exceed those of complex deep learning models, including PointNet++ and Superpoint Graphs, despite using only a random forest and no post-processing.
• The classifier generalizes well across cities, achieving strong performance on Paris after being trained on Lille, even with differing architectural styles and scanning conditions.
• The parameter ρ controls the trade-off between accuracy and speed: mIoU increases sharply up to ρ=3 and plateaus beyond ρ=5, making ρ=5 a practical balance between performance and efficiency.
• The method demonstrates that high-quality handcrafted features can achieve performance comparable to deep learning models, suggesting that feature design remains a valuable direction in 3D semantic segmentation.