Point-GN: A Non-Parametric Network Using Gaussian Positional Encoding for Point Cloud Classification

TL;DR

The paper tackles the challenge of efficient 3D point cloud classification under resource constraints. It introduces Point-GN, a non-parametric framework that combines Gaussian Positional Encoding with non-learnable components such as FPS and k-NN to extract both local and global geometric information without trainable parameters. The core contributions include a four-stage non-parametric feature encoder powered by GPE, a memory-based similarity classifier, and extensive evaluations showing competitive accuracy and fast inference on ModelNet40 and ScanObjectNN, with zero learned parameters. This approach offers a practical solution for real-time perception in robotics and embedded systems, enabling accurate classification without the overhead of parametric models.

Abstract

This paper introduces Point-GN, a novel non-parametric network for efficient and accurate 3D point cloud classification. Unlike conventional deep learning models that rely on a large number of trainable parameters, Point-GN leverages non-learnable components-specifically, Farthest Point Sampling (FPS), k-Nearest Neighbors (k-NN), and Gaussian Positional Encoding (GPE)-to extract both local and global geometric features. This design eliminates the need for additional training while maintaining high performance, making Point-GN particularly suited for real-time, resource-constrained applications. We evaluate Point-GN on two benchmark datasets, ModelNet40 and ScanObjectNN, achieving classification accuracies of 85.29% and 85.89%, respectively, while significantly reducing computational complexity. Point-GN outperforms existing non-parametric methods and matches the performance of fully trained models, all with zero learnable parameters. Our results demonstrate that Point-GN is a promising solution for 3D point cloud classification in practical, real-time environments.

Figures (5)

• Figure 1: Illustration of the proposed non-parametric network for point cloud classification.
• Figure 2: Overview of the architecture of our Non-Parametric Feature Encoder. The figure illustrates the main components of Point-GN, including Gaussian Positional Encoding (GPE), local grouping and feature aggregation. Each stage of the network is designed to efficiently capture spatial relationships within the point cloud without the need for learnable parameters.
• Figure 3: Illustration of the non-parametric classifier pipeline. The test feature is compared with the stored feature embeddings from the training set, and similarity scores are computed to assign the most likely class label based on proximity in feature space.
• Figure 4: Test Speed (samples per second) on ScanObjectNN uy2019revisiting and ModelNet40 wu20153d datasets. The plot compares the inference speed of Point-NN zhang2023parameter and Point-GN on four different datasets. Point-GN shows significant improvements in inference speed across all datasets.
• Figure 5: Ablation study results showing the sensitivity of Point-GN's performance to key hyperparameters: (a) Number of neighbors ($K$), (b) Dimension of Gaussian Positional Encoding (GPE), (c) Number of stages, and (d) Sigma ($\sigma$). We compare the performance of the model on ModelNet40 wu20153d (orange) and ScanObjectNN uy2019revisiting (cyan) datasets, showing both the average and best performances.