.png){kind=icon}
.png){kind=icon}
Return
Optimizing 3D point cloud representations for machine learning: Advances in down-sampling techniques
.png){kind=icon}
PRE
AI
DOI：10.1016/j.neucom.2025.132216.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Down-sampling of 3D point clouds is a preprocessing step that optimizes the trade-off between information preservation and computational efficiency in large, unstructured spatial datasets. However, there is a notable lack of systematic evaluation of these diverse down-sampling techniques and their impact on downstream machine learning tasks. In this paper, we present a comprehensive analysis of recent developments in down-sampling approaches for point cloud processing in intelligent systems. We systematically categorize these methods into task-agnostic and task-oriented approaches, the latter being particularly crucial for tailoring data to specific machine learning objectives. Among them, the fusion-based task-oriented approach represented by REPS (Reconstruction-based Point Cloud Sampling) achieves the best performance of 94.10 % on the classification result on ModelNet40. Through a rigorous examination of their underlying mathematical principles, algorithmic complexities, and statistical properties, we reveal the trade-offs between computational efficiency and feature preservation across different methodologies, which are critical for effective neural network training and performance. Our analysis extends to point cloud representations, applications within machine learning and computer vision, and evaluation metrics, highlighting promising research directions in optimization, statistical learning, and neural architecture design for 3D data. By constructing a robust theoretical framework for 3D data representation and computational efficiency, this work addresses the growing demand for scalable processing methods in large-scale datasets for AI applications and standardizing evaluation benchmarks to ensure fair comparisons of learning algorithms.