International Conference on Computer and Knowledge Engineering

Home / 15th International Conference on Computer and Knowledge Engineering

Class-Aware Balanced Point Cloud Donwsampling for Efficient Large-Scale 3D Scene Understanding

Authors :

Mohammad Yousefipour¹ Marjan Naderan² Morteza Jaderyan³

1- Department of Computer Engineering, Faculty of Engineering, Shahid Chamran University of Ahvaz, Ahvaz, Iran 2- Department of Computer Engineering, Faculty of Engineering, Shahid Chamran University of Ahvaz, Ahvaz, Iran 3- Department of Computer Engineering, Faculty of Engineering, Shahid Chamran University of Ahvaz, Ahvaz, Iran

Keywords :

Point Cloud Downsampling،Class-Aware Sampling،Autonomous Vehicles،Voxel grid

Abstract :

Efficient processing of large-scale point clouds is critical for 3D semantic segmentation, object detection, and scene understanding in autonomous driving and robotics. Traditional downsampling techniques, including voxel grid and random sampling, effectively reduce computational burdens but often fail to preserve the semantic structure and class distributions within highly imbalanced datasets. This can lead to the removal of rare classes, limiting the effectiveness of downstream tasks. In this work, we propose a class-aware voxel-based proportional downsampling algorithm that maintains local class distributions while achieving a global downsampling target. By ensuring each class retains representation within voxels, our approach preserves semantic richness under heavy downsampling rates, making it suitable for efficient training and evaluation pipelines. We validate our method on large-scale datasets, demonstrating its effectiveness in retaining minority classes, preserving geometric structures, and maintaining distributional entropy compared to other downsampling methods. This enables scalable yet semantically faithful point cloud processing for real-world 3D perception systems.