GS-Net: Point cloud sampling with graph neural networks

Xiaolei Chen, Jie Chen, Shoumeng Qiu, Xiangyang Xue, Jian Pu

Abstract: Point clouds are critical for applications like autonomous driving and robotics but face challenges from their massive scale and unstructured nature. To address this, we propose GS-Net, a novel Graph Neural Network(GNN)-based Sampling Network. By leveraging graph-based message passing and jumping knowledge connections, GS-Net effectively captures and fuses multi-scale local geometric structures. Unlike conventional methods neglecting task-specific sampling requirements, GS-Net directly learns task-oriented sampling criteria from downstream models, ensuring the sampled point sets preserve essential geometric features. Extensive experiments on classification, retrieval, and registration tasks across diverse datasets—including ModelNet40, ScanObjectNN, and our newly introduced SemanticKITTI-cls—demonstrate GS-Net’s superior performance compared to existing sampling methods. Remarkably, GS-Net achieves 62.32% classification accuracy on ModelNet40 using only 8 sampled points. Ablation studies confirm GS-Net’s robustness and identify current limitations, providing insights for future work. The source code is available at: https://github.com/chenxl124578/GS-Net.