Prioritising geometric saliency: rethinking point sampling as a discriminative mechanism for 3D point cloud learning

Abstract

Point cloud representation learning heavily relies on input sampling, yet mainstream strategies like random sampling and farthest point sampling (FPS) prioritize spatial coverage over geometric saliency, often discarding discriminative edges and fine structures. To address this, we propose Geometry-Aware Sampling (GAS), a parameter-free preprocessing module that explicitly models point importance through a two-tier geometric prior. First, a Local Multi-scale Saliency (LMS) module quantifies neighborhood spatial dispersion to prioritize edges, corners, and complex regions. Second, a Global Structural Consistency (CSSA) module enforces distributional balance, preventing local oversampling and preserving overall shape integrity. By fusing these scores, GAS constructs an adaptive sampling probability distribution without introducing learnable parameters or altering downstream architectures. Extensive experiments on ModelNet40, ShapeNetPart, and ScanObjectNN demonstrate that GAS consistently enhances both classification and segmentation performance across mainstream networks. Specifically, GAS elevates PointNet++ accuracy from 92.6% to 93.7% and PointTransformer from 93.2% to 94.9% on ModelNet40. Furthermore, GAS maintains strong robustness under noise and sparsity while incurring computational latency comparable to FPS. These results confirm that explicitly integrating geometric priors at the sampling stage significantly improves feature representation quality, offering a highly compatible and reproducible solution for efficient 3D vision pipelines.