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

Marzieh Mohammadi, Amir SalarpourWACV 2025

TL;DR

This paper introduces Point-GN, a novel non-parametric network for 3D point cloud classification that achieves state-of-the-art performance without any trainable parameters. By leveraging Gaussian Positional Encoding (GPE) with existing non-learnable components like Farthest Point Sampling (FPS) and k-Nearest Neighbors (k-NN), Point-GN offers a highly efficient and accurate solution for real-time applications.

What Problem Does This Paper Address?

Classifying 3D point clouds is crucial for many applications, from robotics to medical imaging. However, the irregular and unordered nature of point cloud data makes it challenging for traditional computer vision techniques. While deep learning models like PointNet and PointNet++ have enabled direct processing of point clouds, they often rely on a massive number of trainable parameters, leading to high memory usage and long training times. This complexity hinders their deployment in resource-constrained or real-time scenarios. Existing non-parametric methods often trade off accuracy for efficiency. This paper aims to bridge this gap by developing a non-parametric model that achieves both high accuracy and computational efficiency.

Methods

Point-GN's core innovation lies in its use of Gaussian Positional Encoding (GPE) within a non-parametric framework. Instead of learning positional information, GPE embeds spatial context into point features using Gaussian functions. This allows the network to capture local and global geometric structures without requiring any trainable parameters.

The architecture consists of:

  • Gaussian Positional Encoding (GPE): This module transforms raw 3D coordinates into a higher-dimensional feature space. The encoding is formulated using Gaussian kernels, where the standard deviation ($\sigma$) controls the focus on local versus global spatial information.
  • Non-Parametric Feature Encoder: This hierarchical encoder uses GPE along with standard non-learnable components:
  • Farthest Point Sampling (FPS): Used to downsample the point cloud and select representative local centers.
  • k-Nearest Neighbors (k-NN): Used to find neighbors for each downsampled point.
  • GPE Aggregation: Combines GPE-encoded spatial information with retrieved features, emphasizing closer points.
  • Neighbor Pooling: Aggregates features using mean and max operations. This encoder operates in multiple stages, progressively capturing richer spatial relationships.

  • Non-Parametric Classifier: This classifier uses a similarity-based approach. It compares the feature vector of a test point cloud with pre-computed feature embeddings from the training set. A similarity score is calculated, which is then used to weight label embeddings, ultimately predicting the class with the highest score via a softmax function. This avoids the need for a traditional, learnable classifier.

Datasets

The paper evaluates Point-GN on two widely recognized datasets:

  • ModelNet40: A dataset of clean, synthetic 3D CAD models across 40 object categories. It's used for benchmarking in a controlled environment.
  • ScanObjectNN: A more challenging dataset featuring real-world scenarios with occlusions, background noise, and partial objects. It includes three subsets: OBJ-BG, OBJ-ONLY, and PB-T50-RS, which test robustness under varying complexities.

Key Insights and Ideas

  • Zero Trainable Parameters, High Performance: The most striking insight is that a network can achieve competitive, and in some cases superior, performance in 3D point cloud classification without any learnable parameters. This challenges the conventional wisdom that deep learning models require extensive training.
  • GPE as a Powerful Non-Learnable Feature Extractor: Gaussian Positional Encoding proves to be a highly effective mechanism for embedding spatial information. It allows the network to capture intricate geometric details and relationships that are crucial for accurate classification, without the overhead of learning.
  • Efficiency for Real-Time Applications: By eliminating trainable parameters, Point-GN significantly reduces computational complexity and memory requirements, making it an ideal candidate for real-time applications where resources are limited.
  • Robustness to Real-World Scenarios: Point-GN demonstrates strong performance on the challenging ScanObjectNN dataset, indicating its ability to handle noisy and occluded real-world data effectively.

Key Results

  • ModelNet40: Achieved an accuracy of 85.3%, outperforming the non-parametric Point-NN by +3.5%.
  • ScanObjectNN: Achieved an average accuracy of 85.89% across its splits, outperforming most existing methods, including parametric ones, and significantly outperforming Point-NN. On the challenging PB-T50-RS split, it showed a +21.5% improvement over Point-NN.
  • Inference Speed: Point-GN achieved an inference speed of 301 samples/second on their system, significantly faster than Point-NN.
  • Few-Shot Learning: Point-GN demonstrated superior performance in few-shot classification tasks on ModelNet40, outperforming both parametric and non-parametric methods, highlighting its strong generalization capabilities.

Limitations and Future Work

  • Limitations: While Point-GN achieves impressive results, the paper notes that its accuracy on ModelNet40 (85.3%) is slightly lower than the top-performing parametric models (e.g., PointMLP at 94.1%). However, this comes with a significant trade-off in computational cost and parameter count.
  • Future Work: The authors plan to extend the non-parametric framework by incorporating additional point cloud features. They also aim to explore Point-GN's potential in more complex 3D tasks, such as semantic segmentation and object detection, to further enhance its versatility and real-world applicability.

Summary updated

BibTeX
@inproceedings{mohammadi2025point,
  author = {Marzieh Mohammadi and Amir Salarpour},
  title = {{Point-GN: A Non-Parametric Network using Gaussian Positional Encoding for Point Cloud Classification}},
  booktitle = {IEEE/CVF Winter Conference on Applications of Computer Vision},
  year = {2025},
}
Figure